Prompt injection: Instructions to an AI in a calendar invite

Detects calendar invites whose summary or description addresses an AI (an assistant, a named model such as 'Gemini,', 'the user', 'previous instructions') where NLU finds a request and the prompt injection model agrees on the same field. Calendar assistants read the full event text. Excludes graymail from senders passing DMARC.

Sublime rule (View on GitHub)

 1name: "Prompt injection: Instructions to an AI in a calendar invite"
 2description: "Detects calendar invites whose summary or description addresses an AI (an assistant, a named model such as 'Gemini,', 'the user', 'previous instructions') where NLU finds a request and the prompt injection model agrees on the same field. Calendar assistants read the full event text. Excludes graymail from senders passing DMARC."
 3type: "rule"
 4severity: "low"
 5source: |
 6  type.inbound
 7  //
 8  // This rule makes use of a beta feature and is subject to change without notice
 9  // using the beta feature in custom rules is not suggested until it has been formally released
10  //
11  and any(filter(attachments,
12                 .content_type in ("text/calendar", "application/ics")
13                 or .file_extension =~ "ics"
14          ),
15          any(file.explode(.),
16              any(.scan.ics.calendars,
17                  any(.components,
18                      any([.summary, .description],
19                          regex.icontains(.,
20                                          '\b(?:assistant|llm|screener|summari[sz]er)s?\b|(?:^|[\n\[(>\x27\".:])\s*(?:hey\s+)?(?:chatgpt|gemini|copilot|claude)\s*[:,]|\b(?:ai|language)\s+models?\b|\bthe\s+user\b|\b(?:previous|prior)\s+instructions\b'
21                          )
22                          and any(ml.nlu_classifier(.).entities,
23                                  .name == "request"
24                          )
25                          and beta.prompt_injection_detection(.).detection == "injection"
26                          and beta.prompt_injection_detection(.).confidence in (
27                            "medium",
28                            "high"
29                          )
30                      )
31                  )
32              )
33          )
34  )
35  
36  // graymail from authenticated senders... AI-product newsletters and webinars
37  and not coalesce(ml.attack_score_flash().verdict == "graymail"
38                   and headers.auth_summary.dmarc.pass,
39                   false
40  )  
41tags:
42  - "Attack surface reduction"
43  - "Email prompt injection"
44attack_types:
45  - "BEC/Fraud"
46tactics_and_techniques:
47  - "Evasion"
48  - "Prompt injection"
49  - "Social engineering"
50detection_methods:
51  - "Content analysis"
52  - "File analysis"
53  - "Natural Language Understanding"
54  - "Sender analysis"
55id: "1b4f35e1-91eb-5d6d-8b20-38aff537a1c1"

Related rules

to-top