Detects a single AWS principal invoking Amazon Bedrock models from three or more countries and three or more source IPs within the lookback window. LLMjacking operators often distribute stolen AWS credentials across proxies or resell them to multiple buyers, which produces inference traffic for one identity from many unrelated locations.

Read More

Detects the same AWS principal invoking Amazon Bedrock models (InvokeModel, Converse, and their streaming variants) from geographically distant source IPs within a short window at an impossible travel speed. Stolen access keys and session credentials are frequently resold or reused for LLMjacking, and an inference call from a second continent minutes after a legitimate one is a strong indicator that the credential is being used by someone else.

Read More

Detects the first time in 14 days that a Vertex AI caller (user or service account) runs model inference (GenerateContent, StreamGenerateContent, Predict, or RawPredict) from a source autonomous system (ASN). Stolen Google Cloud credentials, including exported service account keys and OAuth tokens, are used for LLMjacking from attacker infrastructure, which shows up as the victim identity calling models from a hosting provider, VPN, or residential proxy network it has never used before.

Read More

Identifies an IAM policy change on a Google Cloud project, folder, or organization that grants a Vertex AI role (roles/aiplatform.*) to a principal. Vertex AI roles control who can invoke models, deploy endpoints, run agents, and read prompt and training data. An adversary who has gained IAM administration rights can grant a Vertex AI role to an identity they control to keep model access after the original foothold is removed, or to run LLMjacking workloads under the victim's billing. Google-managed service agent roles are excluded because Google grants them automatically when the API is enabled.

Read More

Identifies the creation, modification, or deletion of a content filter (RAI policy) or a custom blocklist on a Microsoft Foundry (Azure AI Services) account. Content filters decide which prompt and completion categories, jailbreak attempts, and protected material a model deployment blocks. An adversary with control of the account can weaken or remove these filters so that harmful prompts and outputs are no longer blocked or logged as filtered, which removes a guardrail and the detections that depend on its verdicts.

Read More

Identifies the first time in seven days that a high-privilege Azure principal lists or regenerates the API keys of a Microsoft Foundry (Azure AI Services) account, or reads the stored secrets of a Foundry account or project connection. Account keys grant direct inference access to every model deployment on the account and bypass Entra ID conditional access, and connection secrets hold credentials for downstream services such as Azure OpenAI, storage, and search. A compromised identity retrieving these keys is a common first step in LLMjacking and in pivoting from a Foundry project to the data it is connected to.

Read More

Detects ChatGPT Enterprise user messages that combine two or more common jailbreak techniques: overriding prior or system instructions, switching the assistant into an unrestricted persona or mode (for example DAN or developer mode), asking it to operate without safety restrictions or filters, or asking it to reveal its system prompt. Jailbreak prompts stack these techniques to talk the model out of its safety policy, and a user doing so in a managed workspace may be attempting to generate prohibited content or to probe the workspace's controls. Requiring more than one technique keeps documents that merely quote a single jailbreak phrase from alerting.

Read More

Identifies a cloud CLI command that destroys infrastructure or identities, such as terminating EC2 instances, removing S3 buckets, or deleting IAM users, when it is spawned by a GenAI coding agent directly or through the agent's shell wrapper. A compromised or prompt-injected agent can be steered into wiping the developer's environment and cloud account using the credentials it already holds, as seen in the malicious Amazon Q Developer for VS Code v1.84.0 release (AWS-2025-015).

Read More

Identifies a Python, Node.js, Bun, Deno, Ruby, PHP or Perl runtime, or a binary executing from a package install location, that opens a developer secret (SSH keys, package registry tokens, source control credentials or AI coding assistant credentials) and then connects to a cloud instance metadata service (IMDS) endpoint within one minute. This sequence matches credential stealers delivered through compromised packages, such as the malicious LiteLLM PyPI releases, while cloud provider credential files are excluded because legitimate cloud SDK credential chains read them before falling back to IMDS.

Read More

Identifies DNS queries to known Large Language Model domains by unsigned binaries or common Windows scripting utilities. Malwares may leverage the capabilities of LLM to perform actions in the affected system in a dynamic way.

Read More

Detects ChatGPT Enterprise audit events where a user role is updated to account-owner. Account owners control workspace membership, connectors, usage limits, and other security-sensitive settings. An adversary who promotes a compromised account to owner gains durable control-plane access.

Read More

Detects ChatGPT Enterprise app authentication events that link a third-party app to a user or workspace. Linking apps establishes durable OAuth-style access that adversaries can use to persist connectors after initial compromise.

Read More

Detects ChatGPT Enterprise assistant messages that recognize likely exposed secrets in chat and advise rotating, revoking, regenerating, invalidating, or deactivating credentials (API keys, access keys, passwords, tokens, and similar). Users or adversaries who paste live secrets into conversations create credential-exposure risk; the assistant's rotation advice is a strong signal that a secret was present in the prior turns.

Read More

Detects successful ChatGPT Enterprise authentication events where the user agent looks scripted (for example Python, curl, Go HTTP clients, aiohttp, scrapers, or headless browsers) rather than a normal browser. Adversaries and automation often reuse stolen sessions or tokens with non-browser clients.

Read More

Detects ChatGPT Enterprise Codex events that create an access token. Codex access tokens enable automated CLI or environment access and can be abused for persistence or lateral automation after account compromise. Treat tokens with a missing expires_at as higher priority during triage.

Read More

Detects ChatGPT Enterprise Codex prompt or response text that matches known credential patterns. Secrets pasted into Codex sessions can be retained in compliance logs and may be reused by agent tool calls.

Read More

Detects ChatGPT Enterprise Codex security-log events that create or update a repository scan configuration. Changing scan scope, lookback, or notification rules can weaken secrets scanning coverage or redirect findings.

Read More

Detects ChatGPT Enterprise conversation messages whose text matches known credential patterns (cloud keys, GitHub tokens, private keys, Slack/SendGrid tokens, Stripe/OpenAI/Anthropic keys, Vault/Docker tokens, JWTs, or Azure storage keys). Secrets pasted into chats can be retained in compliance logs and model context.

Read More

Detects ChatGPT Enterprise audit events where a conversation is shared. Sharing can expose prompts, completions, and attached files outside the original conversation participants and may be used to exfiltrate sensitive content.

Read More

Detects ChatGPT Enterprise custom-agent publish events (agent_published). Publishing an agent with connectors, skills, or store visibility can establish durable automation and broaden data access beyond the original user.

Read More

Detects ChatGPT Enterprise custom-agent events that create or update a trigger. Triggers can schedule or event-drive agent runs without interactive user presence, enabling persistence and automation abuse.

Read More

Detects ChatGPT Enterprise audit events that create a workspace domain boundary entry. Domain changes can expand who is allowed to join or access the tenant and may establish persistence for attacker-controlled domains.

Read More

Detects an unusually high number of ChatGPT Enterprise conversation deletion events for the same user within a rolling 24-hour lookback. Mass deletion can indicate an attempt to remove conversation history or cover tracks after misuse.

Read More

Detects ChatGPT Enterprise Gmail connector send_email responses. Adversaries who abuse a linked Gmail connector after session compromise can exfiltrate data or perform lateral phishing by sending mail on behalf of the user. Draft-only activity (create_draft) is excluded.

Read More

Detects a ChatGPT Enterprise user downloading five or more distinct Google Drive files through the Drive connector within a one-hour window. Adversaries who abuse a linked Drive connector after session compromise often stage bulk file retrieval before sharing, forwarding, or other exfiltration.

Read More

Detects a ChatGPT Enterprise user issuing five or more distinct Google Drive app search queries in a five-minute bucket where at least one query matches sensitive-data themes (payroll, payslip, confidential, payment, NDA, credentials, and similar). Adversaries who abuse a linked Drive connector after session compromise often run rapid multi-term searches to locate high-value files before download or sharing.

Read More

Detects a ChatGPT Enterprise Codex user invoking a high number of distinct MCP tools across multiple connectors within a 30-minute bucket. Broad tool and connector diversity in a short window can indicate compromised-account recon, unauthorized multi-system collection, or an agent rapidly enumerating organizational data sources.

Read More

Detects a high number of ChatGPT Enterprise file download audit events for the same user within a rolling 24-hour lookback. Bulk downloads can indicate staging of conversation attachments or workspace files for exfiltration.

Read More

Detects successful ChatGPT Enterprise logins for the same user email from source IPs whose geo-locations are separated by at least 1,000 km, with implied travel faster than 800 km/h, within a 24-hour window. That pattern can indicate account sharing, VPN or proxy egress mismatches, or an adversary authenticating from a geography far from the legitimate user's baseline.

Read More

Detects ChatGPT Enterprise audit events that delete stored memory. Adversaries may delete memory entries to remove evidence of prior prompts, preferences, or staging instructions after misuse.

Read More

Detects multiple failed ChatGPT Enterprise login attempts for the same user email within a short window. Password spraying or credential stuffing against enterprise accounts often precedes a successful compromise.

Read More

Detects three or more ChatGPT Enterprise assistant messages for the same user whose text matches common policy-refusal phrases (cannot assist, cannot help, cannot fulfill, unable to assist, against my guidelines, and similar) within the lookback window. A refusal burst can indicate jailbreak probing, policy-bypass attempts, or automated unsafe prompting intended to evade model safety controls.

Read More

Detects ChatGPT Enterprise audit events that publish a GPT or update sharing for GPTs, skills, or projects (gpt_publish, update_sharing, skill_share, project_update_sharing). Broadening share scope can expose custom instructions, tools, or proprietary agent configurations beyond the original audience.

Read More

Detects a ChatGPT Enterprise user issuing ten or more distinct Slack app search queries in a five-minute bucket. Adversaries who abuse a linked Slack connector after session compromise often run rapid multi-query searches across channels and messages to locate sensitive content before collection or exfiltration.

Read More

Detects ChatGPT Enterprise Gmail connector searches that look like sensitive discovery: attachment hunts across multiple document types, or queries matching credential, financial, HR, or legal themes. Attackers abusing a linked Gmail connector after session compromise often search mail for high-value content before exfiltration or send-as abuse.

Read More

Detects the first ChatGPT Enterprise Codex MCP tool name used by a given user email within an organization in the rule history window. Codex MCP tools can query external systems and pull organizational data into an agent session. Alerting on first-seen organization, user, and tool triples highlights new or unexpected connector usage rather than every routine approved tool call.

Read More

Detects the first ChatGPT Enterprise Codex plugin install for a given plugin display name and plugin name pair within the rule history window. Newly installed Codex plugins expand the tools and data sources available inside agent sessions; adversaries can enable unexpected connectors to persist access to mail, chat, or file systems after initial compromise.

Read More

Detects the first time a ChatGPT Enterprise MCP app name is used by a given user email within an organization in the rule history window. MCP apps connect ChatGPT to external systems and can move organizational data into third-party tools. Alerting on first-seen organization, user, and app id triples highlights new or unexpected connector usage rather than every routine approved MCP request. App name is still required in the query for investigation context.

Read More

Detects ChatGPT Enterprise activity for the same user email from at least three countries and two unique user agents within a four-hour interval across any compliance data stream. Stolen session cookies often produce post-auth activity (conversation views, deletions, downloads) without a new login event, so this rule is not limited to authentication logs.

Read More

Detects ChatGPT Enterprise audit events that invite a new user into the workspace. An adversary who compromises an administrator can invite a mailbox they control to gain durable access.

Read More

Detects ChatGPT Enterprise audit events that toggle a workspace feature flag (for example GPT sharing). Adversaries may enable sharing or weaken controls to persist access or broaden data exposure.

Read More

Detects ChatGPT Enterprise audit events where a workspace compliance log file is deleted (delete_workspace_log_file). Adversaries with privileged access may delete exported audit or conversation log archives to remove evidence from compliance workflows and hinder investigation.

Read More

Detects ChatGPT Enterprise audit events where workspace compliance log files are listed or downloaded. Attackers pull audit and conversation exports to understand defender visibility or stage sensitive transcript data before further abuse.

Read More

Detects shells or scripting interpreters launched from a parent command line that contains common LLM prompt-injection or jailbreak language. Attackers (or malicious content ingested by an agent) use instruction-override, DAN/developer- mode, system-prompt extraction, and safety-filter bypass phrasing to coerce GenAI coding agents and automation wrappers into running attacker-chosen commands. When that injected text reaches a child shell or interpreter, it is a strong signal that untrusted prompt content is crossing into local code execution, enabling credential theft, data exfiltration, persistence, or further tool abuse under the user's privileges.

Read More

Identifies a burst of failed tool invocations against a single AWS Bedrock AgentCore Gateway within a short window. Individual tool calls fail routinely from bad input or transient upstream errors, but a concentrated run of failures against one gateway can indicate an attacker, or a compromised tool wrapper, probing a tool's parameter surface for an injection point, an authorization bypass, or a schema weakness. Failures can be attributed only to the gateway, not to a specific tool, target, or caller, and a target that is broken for every caller produces the same burst as one caller probing it, so confirm gateway and target health before treating a burst as malicious.

Read More

Detects the successful creation of an Amazon Bedrock AgentCore Browser or Code Interpreter with an execution IAM role and public network access. These sandboxes run agent-generated code or automated browsing on the caller's behalf, and the attached role lets the sandbox call other AWS services. Public network mode removes the sandbox's network containment, so a sandbox steered by prompt injection, malicious tool output, or code-execution abuse can reach the internet and pivot into other AWS resources with the role's permissions. Review the role scope and whether public egress is required.

Read More

Detects the first time a Vertex AI Agent Engine caller executes code in a sandboxEnvironment from a given source autonomous system. First-time sandbox execution by a user and source ASN can indicate a new workload, stolen credentials used from an unusual network, or early Agent Engine abuse before broader sandbox reuse is visible.

Read More

Detects Vertex AI authenticated caller activity for the same client.user.email from at least three countries and three source IPs within a four-hour interval. That pattern can indicate stolen user or service-account credentials reused from multiple regions, including VPN or cloud egress and return trips where first and last locations are close.

Read More

Detects the same Vertex AI authenticated caller (client.user.email) generating aiplatform audit activity from geographically distant source IPs within a short window at an impossible travel speed. That pattern can indicate a stolen user or service-account credential used from multiple regions. Related: GCP Vertex AI Caller Activity From High Number of Countries covers slower multi-country reuse without a high-speed hop.

Read More

Detects Vertex AI prompt-response logs where a candidate safety_ratings.probability is HIGH or MEDIUM. Those values are stronger signals than NEGLIGIBLE or LOW alone. Requires the request to include safetySettings so ratings are logged.

Read More

Detects a single source IP making a high number of Vertex AI GenerateContent (or StreamGenerateContent) audit calls against the same model resource in the lookback window. That pattern fits model extraction, automated scraping, or a compromised caller burning prediction quota.

Read More

Detects a high number of Vertex AI prompt-response log rows and high total token usage for one model in the lookback window. Totals use provider usage_metadata.total_token_count (includes thinking tokens when billed). High request and token volume can indicate model extraction, quota abuse, or a compromised caller. Tune thresholds for your baseline.

Read More

Detects Vertex AI GenerateContent prompts that match common jailbreak / instruction-override heuristics (for example ignore previous instructions, DAN or developer mode, unrestricted-AI language, or explicit safety-filter bypass). Typical prompt_response_logs exports do not populate HARM_CATEGORY_JAILBREAK (unlike hate/harassment/dangerous ratings); Model Armor findings remain the preferred first-party jailbreak signal when that API is enabled.

Read More

Detects a GCP Vertex AI GenerateContent exchange whose prompt or model reply contains a known credential pattern (AWS access keys, GitHub tokens, PEM private keys, and similar), or whose assistant reply warns about exposed keys and revocation. Secrets in prompt/response logs are visible to anyone with access to BigQuery prompt-response export and Elastic.

Read More

Detects changes to Vertex AI publisher model configuration via SetPublisherModelConfig (for example enabling or retargeting prompt/response BigQuery logging). Unexpected config changes can disable security-relevant logging or redirect logs.

Read More

Detects three or more Vertex AI GenerateContent replies in 60 minutes whose assistant text indicates a refusal (cannot fulfill/assist/help/answer, unable to assist, decline the request, against my guidelines, and similar), aggregated by project and model. Prompt-response logs lack a stable caller identity, so this is a project-wide refusal burst signal — pivot to auditlogs for principal/IP attribution.

Read More

Detects Vertex AI GenerateContent responses whose candidate finish_reason is SAFETY. That value is emitted when Gemini safety filters stop generation. Unlike textual refusals (finish_reason STOP), SAFETY means the provider blocked or truncated the completion.

Read More

Detects Vertex AI GenerateContent requests that set safety_settings.threshold to BLOCK_NONE. That threshold disables category blocking for the listed harm categories when clients explicitly send safetySettings. Attackers and careless apps use BLOCK_NONE to weaken built-in Gemini filters.

Read More

The organization admin role controls organization settings, integrations, membership, and security configuration in Anthropic Claude for Enterprise. Membership role changes are reported as claude_user_role_updated with anthropic.audit.current_role. An attacker can promote a compromised or newly invited account to org admin to turn initial access into durable control-plane access. From admin, they can disable SSO, mint admin API keys for automation, start data exports, and weaken audit logging. Workspace-scoped role_assignment_granted grants (for example bare admin on a workspace) are out of scope for this rule.

Read More

Primary ownership is the highest administrative authority in an Anthropic organization, covering billing, membership, and organization-wide settings. Transferring ownership to an attacker-controlled account can lock out the legitimate administrator from recovery paths that depend on the original owner. Attackers often do this after role escalation so defenders cannot reverse earlier privilege changes through normal administration.

Read More

Identifies when access to an Amazon Bedrock foundation model is enabled at the account level, either by granting a foundation-model entitlement, submitting a use case for model access, or creating a foundation-model agreement (accepting the EULA). These account-level "model access" actions unlock a foundation model so that it can subsequently be invoked. Adversaries or a compromised principal may enable model access to abuse expensive models (LLMjacking), to establish a durable ability to invoke models within the account, or to bypass organizational controls. This activity is distinct from changes to a resource-based model invocation policy and is identified by the Bedrock control-plane API calls that grant model entitlements and agreements.

Read More

Detects when an AWS principal using long-term IAM user credentials (AKIA* access key) enumerates available Bedrock foundation models and then invokes a model within the same 15-minute window. Most legitimate Bedrock workloads run under IAM roles with short-lived credentials; the combination of model enumeration followed by direct model invocation from a long-term IAM user key is unusual in production environments and consistent with an adversary using stolen credentials to discover and exploit available AI model capabilities. This pattern is associated with LLMjacking attacks where threat actors abuse compromised cloud credentials to run high-volume or high-cost model inference at the account owner's expense.

Read More

Identifies failed, access-denied attempts to enable account-level access to an Amazon Bedrock foundation model, either by granting a foundation-model entitlement, submitting a use case for model access, or creating a foundation-model agreement (accepting the EULA). These account-level "model access" actions unlock a foundation model so that it can subsequently be invoked. A principal that is repeatedly denied when attempting these actions may be a compromised or under-privileged identity probing for the ability to unlock expensive models (LLMjacking) or to establish a durable ability to invoke models. Unlike the companion rule that detects successful model-access grants, this rule surfaces the attempt itself, which is a high-signal indicator of credential boundary-testing even though access was not granted.

Read More

Detects when Azure OpenAI requests result in zero response length, potentially indicating issues in output handling that might lead to security exploits such as data leaks or code execution. This can occur in cases where the API fails to handle outputs correctly under certain input conditions.

Read More

Monitors for suspicious activities that may indicate theft or unauthorized duplication of machine learning (ML) models, such as unauthorized API calls, atypical access patterns, or large data transfers that are unusual during model interactions.

Read More

Detects patterns indicative of Denial-of-Service (DoS) attacks on machine learning (ML) models, focusing on unusually high volume and frequency of requests or patterns of requests that are known to cause performance degradation or service disruption, such as large input sizes or rapid API calls.

Read More

Detects unusual modification of GenAI tool configuration files. Adversaries may inject malicious MCP server configurations to hijack AI agents for persistence, C2, or data exfiltration. Attack vectors include malware or scripts directly poisoning config files, supply chain attacks via compromised dependencies, and prompt injection attacks that abuse the GenAI tool itself to modify its own configuration. Unauthorized MCP servers added to these configs execute arbitrary commands when the AI tool is next invoked.

Read More

Detects one caller IP sending a high number of requests and consuming a high number of tokens against the same Foundry deployment and operation during the rule lookback. That rate fits a script or a stolen key burning quota. The rule uses RequestResponse logs and does not need the prompt text. Tune the request count and token sum in the query.

Read More

Detects a Microsoft Foundry chat, sent through API Management, where Azure sets jailbreak.detected on the prompt. The flag is on the prompt filter when the call is allowed through, and on the error body when the filter blocks the call.

Read More

Detects a Microsoft Foundry chat, sent through API Management, whose prompt or assistant reply contains a known credential pattern or an email address together with a password assignment. The model often refuses the request and Azure content filters stay clear, so the secret is only visible in the logged message text.

Read More

Detects three or more completed assistant refusals from the same client, API Management subscription, and model. Matching replies finished on their own (finish_reason stop) and declined the prompt. A burst means that caller kept sending prompts the model will not answer.

Read More

Detects the same API Management subscription key calling Microsoft Foundry from source IPs in two or more countries, far enough apart and fast enough that the hop is not physical travel. Calls are grouped by apim_subscription_id. Requests with no subscription key are ignored, because those are unauthenticated calls and share no key.

Read More

Detects Anthropic activity for the same user email from at least three countries and two unique user agents within a four-hour interval. This highly unusual activity may indicate leaked session cookies being abused by a remote adversary using a VPN.

Read More

Detects successful Anthropic audit activity where the user agent matches scripting HTTP clients, offensive scanners, or automation libraries (for example curl, python-requests, Go-http-client, axios, nuclei). Browser and first-party Claude clients normally present recognizable browser or product user agents; raw HTTP library agents on successful control-plane or product activity often indicate scripted access, stolen-session reuse, or unauthorized automation. Known benign automation patterns (axios or Go HTTP clients on file upload/delete, trusted device enrollment, and environment creation) are excluded.

Read More

Admin API keys grant programmatic access to organization and compliance APIs outside an interactive browser session. An attacker who creates one after compromise can automate role grants, exports, and logging changes without holding a user session that would time out under SSO. The key also survives password resets and IdP lockout if defenders revoke the interactive account but miss the API credential.

Read More

Admin API keys grant programmatic access to organization and compliance APIs. An attacker can delete legitimate admin API keys to break security monitoring or integrations, or to cover tracks after creating replacement credentials they control. Deletion without a nearby rotation event points more at sabotage than routine key hygiene.

Read More

Claude artifacts can be shared with specific audiences. Making an artifact public exposes its contents to unauthenticated viewers on the internet. An attacker with access to sensitive artifacts can publish them publicly to push intellectual property, credentials embedded in prompts, or other confidential material outside organizational controls.

Read More

Compliance-scoped API keys read organization audit activity and compliance data. Once an attacker has administrative access, creating one gives them programmatic read of chats, files, and membership without an interactive session. This is separate from admin API key creation, which covers organization administration rather than compliance read scopes.

Read More

Compliance API logging feeds the anthropic.audit dataset that Anthropic audit detections run on. An attacker with administrative access can disable it so later role grants, API key creation, exports, and authentication changes stop reaching this data source. This rule detects the disable action itself. Activity that happens after logging stops may not appear in logs-anthropic.audit-*.

Read More

An audit log export archive was accessed, meaning the actor downloaded exported audit activity. Attackers pull audit exports to see what defenders can observe, look for detection gaps, or remove evidence before making other control-plane changes.

Read More

Detects a single authenticated user generating an unusually high number of denied Claude chat access attempts in a 24-hour window. That pattern fits automated chat enumeration or attempts to reach conversations outside the actor's permissions. Unauthenticated shared-link actors lack user.id and are excluded.

Read More

Detects an unusually high number of Claude chat creation events for the same user email within a 24-hour period. Burst chat creation can indicate automated LLM abuse, resource hijacking to burn organizational quotas, or scripted workflows used to stage many parallel conversations for data processing or prompt-injection campaigns.

Read More

Detects an unusually high number of Claude chat deletion events for the same user email within a single calendar day. Mass chat deletion can indicate an attempt to remove conversation history, cover tracks after misuse of Claude, or automated cleanup following data staging or prompt abuse.

Read More

Detects an unusually high number of successful Claude chat snapshot creation events for the same user email within a rolling 24-hour window. Chat snapshots package conversation content into shareable exports; sustained creation volume can indicate staging of organizational chat data for exfiltration via Anthropic's web service or automated bulk export of sensitive prompts and responses.

Read More

Extra usage spend limits cap Anthropic organizational spend beyond included usage. Deleting a spend limit removes that cap and can enable unrestricted API or Claude consumption. An attacker who already has administrative or API access can delete the limit to burn budget, run large automated workloads, or stage resource abuse without the previous guardrail.

Read More

Detects an unusually high volume of Claude file uploads from the same user email and source IP within a rolling 24-hour period. Sustained upload activity can indicate staging of sensitive documents in Claude chats for later retrieval, automated ingestion of data into LLM workflows, or abuse of organizational Claude access to move files into the cloud service.

Read More

Detects successful Anthropic magic link or SSO sign-ins for the same user email from source IP addresses whose query-time geo-locations (via IP_LOCATION) are separated by at least 1,000 km, with implied travel faster than 800 km/h, within a 24-hour window. That pattern can indicate account sharing, VPN or proxy egress mismatches, or an adversary authenticating from a geography far from the legitimate user's baseline.

Read More

Magic link second factor adds an extra authentication step to passwordless sign-in for Anthropic. An attacker with administrative access can turn it off so magic link logins no longer require the second factor, which makes stolen or attacker-controlled mailboxes usable for interactive access. This often shows up alongside SSO weakening when the attacker wants a fallback authentication path outside the corporate IdP.

Read More

Detects the first successful creation of a Model Context Protocol (MCP) server integration name in an Anthropic organization within the rule history window. MCP servers add external data pathways into Claude and can expose organizational data to third-party infrastructure. This is a New Terms rule keyed on organization.id and anthropic.audit.mcp_server_name so the same connector name can still alert in another tenant, while routine re-creation of an already-seen name in the same organization does not.

Read More

Detects at least five failed Anthropic authentication events for the same user email within one hour. Failures are matched by authentication category and failure outcome (for example magic-link or SSO login failures). That pattern fits repeated guessing, stale magic link abuse, or automated login attempts against one account.

Read More

Starting an organization data export only signals intent. Accessing the export archive via its signed URL means the actor actually downloaded chats, projects, user metadata, and configuration. An attacker with administrative access can use this to exfiltrate intellectual property and credentials at scale.

Read More

Organization deletion and bulk delete remove tenant data, projects, and member access in a single administrative action. An attacker can use this to break AI-assisted workflows as an impact technique, to extort the organization, or to destroy evidence after finishing a data export. Once deletion progresses, recovery options shrink and earlier exfiltration activity is harder to reconstruct from the tenant itself.

Read More

Verified organization domains control which email addresses can join or be pulled into an Anthropic tenant. Verifying, claiming, adding, or removing a domain changes the tenant boundary and can pull in attacker-controlled mailboxes or push out legitimate corporate domains. These events are infrequent and affect organization-wide membership trust.

Read More

Organization IP restrictions limit Anthropic administrative access to approved network ranges. Deleting one widens where a compromised admin session or API key can be used. The audit event does not always carry the deleted CIDR or restriction identifier, so treat this as an early signal and pivot to nearby IP restriction create or update events for the same organization.

Read More

Detects a single user performing at least two distinct organization discovery actions within a 10-minute window: listing users, exporting members, or viewing groups. Chaining these read actions maps membership and group structure and commonly precedes targeted role grants, invites, or data collection against high-value accounts.

Read More

Sending an organization user invite creates a path for a new member to join the Anthropic tenant with a chosen role. An adversary who compromises an administrator or admin API key can invite a mailbox they control and accept the invite to gain durable access. Invites may target internal corporate addresses or external domains; this rule does not distinguish them because invite events do not carry verified organization domains for reliable comparison.

Read More

Detects when a Claude project owner or editor role is granted through a role_assignment_granted event. Project owners and editors can access project chats, artifacts, and knowledge bases that may hold sensitive data. An attacker with organization access can grant these roles to persist access to high-value project content without holding organization admin privileges.

Read More

Detects successful Anthropic audit activity for the same user email from source IP addresses whose query-time geo-locations (via IP_LOCATION) span at least two countries, are separated by at least 500 km, and imply travel faster than 800 km/h within a short (~15-minute) lookback. Unlike login-only impossible travel, this rule covers any successful user-actor activity and can surface session cookie replay or concurrent session reuse when no new authentication events appear.

Read More

SSO routes Anthropic authentication through the corporate identity provider. Disabling SSO, or deactivating or deleting an SSO connection, moves users onto alternate sign-in paths where IdP-enforced MFA, conditional access, and session policies no longer apply. That opens the door to password or magic-link accounts the attacker controls.

Read More

Identifies AWS Bedrock Agent creation performed directly by an IAM user or the root account. Bedrock Agents are autonomous AI systems that execute multi-step tasks, invoke Lambda action groups to call external APIs, and query knowledge bases. Adversaries with access to an AWS account can create rogue agents configured to exfiltrate data via action group Lambda functions, pivot to other services, or act as a persistent AI-driven command-and-control channel. This rule is scoped to IAMUser and Root identity types — AssumedRole sessions (which represent automated CI/CD pipelines and SSO-federated engineers) are excluded to avoid global false positives from legitimate deployment automation that varies widely across customer environments.

Read More

Detects modification of deployed Amazon Bedrock agents and their action groups, collaborators, or aliases via the Bedrock Agent control plane. Adversaries with access to an AWS account can tamper with an existing, trusted agent by altering its instructions (UpdateAgent), adding or changing action groups that wire the agent to Lambda functions or APIs (CreateAgentActionGroup, UpdateAgentActionGroup), attaching or modifying collaborators (AssociateAgentCollaborator, UpdateAgentCollaborator), or repointing an alias to a tampered version (CreateAgentAlias, UpdateAgentAlias). A PrepareAgent call is required to make a tampered configuration live. By implanting malicious behavior into an agent that legitimate users continue to invoke, an attacker can maintain durable access through a trusted component. Creation of brand-new agents (CreateAgent) is intentionally excluded as lower-signal activity.

Read More

Identifies an Amazon Bedrock AgentCore execution role (an AssumedRole identity whose role name begins with "AgentCore-" or contains "BedrockAgentCore") making an AWS API call to a service it has not previously called. AgentCore runtimes normally interact only with Bedrock inference, AgentCore data-plane, and observability services (CloudWatch Logs, X-Ray, CloudWatch metrics), so an execution role suddenly calling STS, EC2, IAM, Secrets Manager, or other services is a strong indicator that the role's temporary credentials were exfiltrated from the agent's microVM (for example, via the Code Interpreter instance-metadata-service credential theft) and are being used outside the runtime for reconnaissance, privilege escalation, or lateral movement. Because the stolen credentials are recorded in CloudTrail under the execution role's own identity, the anomalous service usage, not the identity, is the detectable signal.

Read More

Detects the creation of an AWS Bedrock AgentCore resource (code interpreter, agent runtime, browser, or harness) with an IAM execution role attached. When an attacker with iam:PassRole permission creates an AgentCore resource and attaches a privileged role, subsequent invocations inside that resource execute as the attached role — enabling privilege escalation to roles that trust bedrock-agentcore.amazonaws.com.

Read More

Identifies prompts sent to an Amazon Bedrock AgentCore runtime that contain AWS access key identifiers (AKIA long-term or ASIA temporary/STS), Amazon Bedrock API keys (ABSK bearer tokens), or PEM-encoded private keys. The runtime application logs record the caller-supplied prompt; credentials embedded in a prompt are exposed to the model provider, persisted in observability logs, and may be returned in completions or used by downstream tools. This commonly indicates accidental secret leakage by a user or application, or an attempt to stage credentials for misuse through the agent. Secrets should never be passed to an agent in clear text.

Read More

Identifies prompts sent to an Amazon Bedrock AgentCore runtime that attempt to harvest credentials or coerce the agent into exfiltrating data. The runtime application logs capture the caller-supplied prompt; this rule flags prompts that reference the cloud instance metadata service (169.254.169.254, the ECS task metadata address, or the "latest/meta-data" / "security-credentials" paths), prompts that name AWS access or secret keys directly, and prompt-injection or jailbreak language ("ignore previous instructions", "developer mode", "do anything now") combined with intent to reveal secrets, system prompts, or send data to an external endpoint. Asking an agent to read instance metadata credentials or to exfiltrate secrets is rarely legitimate and indicates an attempt to weaponize the agent for credential theft, even when the model refuses the request.

Read More

Identifies an Amazon Bedrock API key phantom user (an IAM user whose name starts with "BedrockAPIKey-") acting as the caller of a non-Bedrock API request, such as IAM, STS, EC2, VPC, or KMS calls. These users are provisioned by AWS to back a Bedrock bearer token and carry the AmazonBedrockLimitedAccess managed policy, which also grants IAM, VPC, and KMS reconnaissance. A phantom user performing activity outside of Bedrock indicates its credentials are being used beyond their intended scope, which is the privilege-escalation path realized: an attacker who created standard IAM access keys for the phantom user is now using them for reconnaissance or lateral movement outside the Bedrock authentication boundary.

Read More

Identifies an Amazon Bedrock API key (bearer token) being used to perform a destructive or anti-recovery control-plane action, such as deleting a guardrail, deleting a custom or imported model, removing provisioned throughput, or disabling model invocation logging. Bedrock API keys are bearer credentials intended for model invocation (InvokeModel, Converse); using one to delete Bedrock resources or disable logging is inconsistent with that purpose and is characteristic of LLMjacking or sabotage following key theft. Every Bedrock API key call is identifiable in CloudTrail by "additionalEventData.callWithBearerToken" being true. The rule matches regardless of outcome, because a destructive attempt via a bearer token is suspicious even when denied.

Read More

Detects deletion or modification of AWS Bedrock Automated Reasoning policies via the DeleteAutomatedReasoningPolicy, UpdateAutomatedReasoningPolicy, or UpdateAutomatedReasoningPolicyAnnotations CloudTrail actions. Automated Reasoning policies are a Bedrock safety and validation control that constrains model outputs against formal rules. An adversary who deletes a policy or alters the policy definition or its annotations weakens an enforced output-validation defense, potentially allowing unsafe or non-compliant model responses to pass unchecked. Benign build, test-workflow, and test-case CRUD operations are intentionally excluded as they have no coherent abuse path.

Read More

Identifies multiple successive failed attempts to use denied model resources within AWS Bedrock. This could indicated attempts to bypass limitations of other approved models, or to force an impact on the environment by incurring exhorbitant costs.

Read More

Identifies multiple validation exeception errors within AWS Bedrock. Validation errors occur when you run the InvokeModel or InvokeModelWithResponseStream APIs on a foundation model that uses an incorrect inference parameter or corresponding value. These errors also occur when you use an inference parameter for one model with a model that doesn't have the same API parameter. This could indicate attempts to bypass limitations of other approved models, or to force an impact on the environment by incurring exhorbitant costs.

Read More

Detects deletion, weakening, or version management of AWS Bedrock guardrails via the DeleteGuardrail, UpdateGuardrail, DeleteEnforcedGuardrailConfiguration, or PutEnforcedGuardrailConfiguration APIs. Bedrock guardrails enforce content, topic, word, and sensitive-information policies on model invocations. Deleting a guardrail, loosening its policies, removing or overwriting the organization-enforced guardrail configuration, or creating a new version to enforce a weakened configuration allows an adversary to bypass these protections — the cloud control-plane equivalent of disabling a security tool. This activity should be validated against approved change management and the responsible identity.

Read More

Identifies multiple violations of AWS Bedrock guardrails within a single request, resulting in a block action, increasing the likelihood of malicious intent. Multiple violations implies that a user may be intentionally attempting to cirvumvent security controls, access sensitive information, or possibly exploit a vulnerability in the system.

Read More

Identifies multiple violations of AWS Bedrock guardrails by the same user in the same account over a session. Multiple violations implies that a user may be intentionally attempting to cirvumvent security controls, access sensitive information, or possibly exploit a vulnerability in the system.

Read More

Detects when a Bedrock model is prompted to invoke high-risk tools associated with shell execution, filesystem operations, or process spawning. Adversaries may use compromised AI agent pipelines or manipulated prompts to instruct the model to execute arbitrary system commands, read or write sensitive files, or spawn subprocesses — extending the blast radius of a credential compromise or prompt injection attack.

Read More

Identifies an AWS principal performing a high volume of Amazon Bedrock inference API calls against a single model within a short window. Membership inference attacks require hundreds to thousands of statistically similar queries whose prompts and responses are intentionally content-benign, making guardrail- and content-based rules ineffective. This rule detects the high-frequency single-model probing pattern that precedes membership inference and related exfiltration via the inference API. It is a behavioral / volumetric precursor: it does not observe model confidence scores and a fixed call-count threshold only catches the loud variant, so paced, low-and-slow, or credential-distributed probing will evade it. Definitive membership inference detection requires ML anomaly analysis over per-entity inference-rate and response-distribution baselines.

Read More

Identifies multiple AWS Bedrock executions in a one minute time window without guardrails by the same user in the same account over a session. Multiple consecutive executions implies that a user may be intentionally attempting to bypass security controls, by not routing the requests with the desired guardrail configuration in order to access sensitive information, or possibly exploit a vulnerability in the system.

Read More

Detects control-plane mutations to AWS Bedrock knowledge bases and their backing RAG data sources via CloudTrail. An adversary with access to Bedrock Agent APIs can poison the corpus that RAG-enabled models treat as authoritative by ingesting attacker-controlled documents (IngestKnowledgeBaseDocuments, StartIngestionJob), deleting legitimate documents (DeleteKnowledgeBaseDocuments), or repointing/altering the data source itself (CreateDataSource, UpdateDataSource, DeleteDataSource, UpdateKnowledgeBase). Because downstream applications and users trust model answers grounded in this stored data, tampering with the corpus is a stored data manipulation that can drive misinformation, fraud, or manipulated decisions at inference time. This is a New Terms rule that looks for the first time a given identity ARN performs one of these knowledge base or data source mutations within the history window.

Read More

Detects when an AWS Bedrock model invocation logging configuration is deleted or overwritten via the DeleteModelInvocationLoggingConfiguration or PutModelInvocationLoggingConfiguration API calls. Model invocation logging is the source that feeds the logs-aws_bedrock.invocation-* dataset relied upon by all data-plane Bedrock detections. An adversary who has gained access to a Bedrock environment can blind defenders by deleting this configuration, or by using the Put API to redirect logs to an attacker-controlled or non-monitored S3 bucket or CloudWatch log group. Because this single control-plane action can neutralize the entire data-plane detection stack, it is a high-value evasion technique that should be validated against expected administrative change activity.

Read More

Identifies an Amazon Bedrock model invocation whose prompt or completion contains an AWS access key identifier (AKIA long-term or ASIA temporary/STS, followed by 16 characters), an Amazon Bedrock API key (ABSK bearer token), or a PEM private-key block. Credentials in the model input mean an application or user is sending secrets to the model, exposing them to invocation logging, the model provider, and prompt history; credentials in the model output mean the model is emitting secrets, which can result from training-data leakage, poisoned context, or a prompt-injection-driven exfiltration attempt. Either case is a credential-exposure event that warrants immediate rotation of the affected secret.

Read More

Detects creation, modification, or deletion of AWS Bedrock Provisioned Model Throughput via the CreateProvisionedModelThroughput, UpdateProvisionedModelThroughput, and DeleteProvisionedModelThroughput APIs. Provisioned Throughput reserves dedicated, billed model capacity for Amazon Bedrock. An adversary who scales this capacity up can drive large, unauthorized cost (cloud resource/bill hijacking), while deleting reserved throughput can cause denial of service to production workloads that depend on that committed capacity. These control-plane changes should be validated against approved capacity-planning and change-management processes.

Read More

Detects modification or deletion of resource-based access policies on AWS Bedrock resources via the PutResourcePolicy and DeleteResourcePolicy API calls. Resource-based policies govern which principals (including external accounts) may access Bedrock resources such as agents, knowledge bases, and custom models. An adversary may attach a resource policy granting an external or unexpected principal access to a Bedrock resource to establish persistence or enable cross-account access, or may delete an existing policy to weaken access controls. These changes should be validated for principal ownership and least-privilege intent.

Read More

Detects when an Amazon Bedrock agent is associated with, or updated to use, a knowledge base via the AssociateAgentKnowledgeBase, or UpdateAgentKnowledgeBase API actions. Bedrock agents consume knowledge base (RAG) content as trusted context for the model. By wiring an agent to an externally controlled or third-party knowledge base, or by swapping in an attacker-controlled knowledge base, an adversary can redraw the agent's trust boundary toward an untrusted source. This is a software-supply-chain compromise and an indirect prompt-injection delivery vector: poisoned or adversarial content served from the associated knowledge base is treated as authoritative by the agent. Validate that the associated knowledge base, and any underlying data source, is owned and controlled by your organization.

Read More

Detects failed, access-denied attempts to modify or delete resource-based access policies on AWS Bedrock resources via the PutResourcePolicy and DeleteResourcePolicy API calls. Resource-based policies govern which principals (including external accounts) may access Bedrock resources such as agents, knowledge bases, and custom models. A principal that is repeatedly denied when attempting to attach or remove these policies may be a compromised or under-privileged identity probing for the ability to grant external or cross-account access, or to weaken existing access controls. Unlike the companion rule that detects successful changes, this rule surfaces the attempt itself, which is a high-signal indicator of credential boundary-testing even though no change occurred.

Read More

Detects when an AWS Bedrock custom model is imported or deployed, or when a marketplace model endpoint is created or registered, via the CreateModelImportJob, CreateCustomModelDeployment, CreateMarketplaceModelEndpoint, or RegisterMarketplaceModelEndpoint API calls. These actions introduce a model artifact from outside the organization's trusted training and approval pipeline. A backdoored, poisoned, or attacker-supplied model that downstream applications subsequently invoke represents a software supply-chain compromise. New model imports and marketplace endpoint registrations should be validated for artifact provenance (S3 source ownership), the registering identity, and whether the model originates from an approved internal pipeline.

Read More

Identifies standard IAM credentials being added to an Amazon Bedrock API key phantom user, whose user name starts with "BedrockAPIKey-": either a long-term access key (CreateAccessKey) or a console password / login profile (CreateLoginProfile, UpdateLoginProfile). When a long-term Bedrock API key is generated through the AWS Console, AWS silently provisions a "BedrockAPIKey-" IAM user with the AmazonBedrockLimitedAccess managed policy. That user is intended only to back a Bedrock bearer token and should never hold standard programmatic keys or interactive console access. Adding either converts a Bedrock-scoped identity into general-purpose IAM credentials that inherit the policy's Bedrock control-plane and IAM, VPC, and KMS reconnaissance permissions and that persist after the Bedrock API key is revoked. This is the privilege-escalation and persistence pivot documented for Bedrock API key phantom users, and there is no legitimate workflow that produces it.

Read More

Detects unexpected modification of Claude Desktop Cowork VM boot images (kernel, initrd, root filesystem). Adversaries with user-context access can rewrite these stored images so later Cowork sessions boot attacker-controlled code inside a virtual instance that host EDR cannot inspect by default.

Read More

Detects Elastic Defend alerts (behavior, malicious file, memory signature, shellcode) where the alerted process or its direct parent is a GenAI coding or assistant utility (e.g. Cursor, Claude, Windsurf, Cody, Continue, Aider, OpenClaw, Moltbot, Clawdbot, Codeium, Tabnine, GitHub Copilot). Activity from these tools can indicate prompt injection, malicious skills, or supply-chain abuse; this Higher-Order rule helps prioritize such alerts for triage.

Read More

Detects suspicious child process execution from the OpenClaw, Moltbot, or Clawdbot AI coding agents running via Node.js. These tools can execute arbitrary shell commands through skills or prompt injection attacks. Malicious skills from public registries like ClawHub have been observed executing obfuscated download-and-execute commands targeting cryptocurrency wallets and credentials. This rule identifies shells, scripting interpreters, and common LOLBins spawned by these AI agents.

Read More

Identifies GenAI agent CLIs started with permission-bypass or auto-approval flags that disable human-in-the-loop guardrails. These modes are intended for isolated sandboxes but are frequently misused on internet-connected developer workstations, allowing prompt injection, compromised dependencies, or malicious skills to execute commands, modify files, or reach sensitive paths without confirmation.

Read More

Detects when GenAI tools access sensitive files such as cloud credentials, SSH keys, browser password databases, or shell configurations. Attackers leverage GenAI agents to systematically locate and exfiltrate credentials, API keys, and tokens. Access to credential stores (.aws/credentials, .ssh/id_*) suggests harvesting, while writes to shell configs (.bashrc, .zshrc) indicate persistence attempts. Note: On linux only creation events are available. Access events are not yet implemented.

Read More

Detects when GenAI tools spawn compilers or packaging tools to generate executables. Attackers leverage local LLMs to autonomously generate and compile malware, droppers, or implants. Python packaging tools (pyinstaller, nuitka, pyarmor) are particularly high-risk as they create standalone executables that can be deployed without dependencies. This rule focuses on compilation activity that produces output binaries, filtering out inspection-only operations.

Read More

Detects when GenAI tools connect to domains using suspicious TLDs commonly abused for malware C2 infrastructure. TLDs like .top, .xyz, .ml, .cf, .onion are frequently used in phishing and malware campaigns. Legitimate GenAI services use well-established domains (.com, .ai, .io), so connections to suspicious TLDs may indicate compromised tools, malicious plugins, or AI-generated code connecting to attacker infrastructure.

Read More

Detects GenAI tools connecting to unusual domains on macOS. Adversaries may compromise GenAI tools through prompt injection, malicious MCP servers, or poisoned plugins to establish C2 channels or exfiltrate sensitive data to attacker-controlled infrastructure. AI agents with network access can be manipulated to beacon to external servers, download malicious payloads, or transmit harvested credentials and documents.

Read More

Detects when GenAI processes perform encoding or chunking (base64, gzip, tar, zip) followed by outbound network activity. This sequence indicates data preparation for exfiltration. Attackers encode or compress sensitive data before transmission to obfuscate contents and evade detection. Legitimate GenAI workflows rarely encode data before network communications.

Read More

Detects when the Ollama LLM server accepts connections from external IP addresses. Ollama lacks built-in authentication, so exposed instances allow unauthenticated model theft, prompt injection, and resource hijacking.

Read More

Detects potential resource exhaustion or data breach attempts by monitoring for users who consistently generate high input token counts, submit numerous requests, and receive large responses. This behavior could indicate an attempt to overload the system or extract an unusually large amount of data, possibly revealing sensitive information or causing service disruptions.

Read More

Detects repeated high-confidence 'BLOCKED' actions coupled with specific 'Content Filter' policy violation having codes such as 'MISCONDUCT', 'HATE', 'SEXUAL', INSULTS', 'PROMPT_ATTACK', 'VIOLENCE' indicating persistent misuse or attempts to probe the model's ethical boundaries.

Read More

Detects repeated compliance violation 'BLOCKED' actions coupled with specific policy name such as 'sensitive_information_policy', indicating persistent misuse or attempts to probe the model's denied topics.

Read More

Detects repeated compliance violation 'BLOCKED' actions coupled with specific policy name such as 'topic_policy', indicating persistent misuse or attempts to probe the model's denied topics.

Read More

Detects repeated compliance violation 'BLOCKED' actions coupled with specific policy name such as 'word_policy', indicating persistent misuse or attempts to probe the model's denied topics.

Read More