T01 · Skill Instruction Hijacking
- Location
scripts/semantic_search.py:185- Finding
Untrusted Email Content Is Passed to the Agent Without a Prompt-Injection Boundary
- Content
View full analysis
limit, "emails": emails } print(json.dumps(result, ensure_ascii=False)) ``` ### Technical Analysis Email subjects, sender and recipient fields, and body text are externally controlled. Any Internet user who can send mail to the account can place natural-language instructions in these fields. The semantic-search implementation extracts up to 500 characters from each email body and places that content directly in the tool result consumed by the Agent. The Skill instructions state that this result is supplied to the Agent for semantic analysis, but they do not establish that all email content is untrusted data or prohibit the Agent from following instructions contained in messages. Removing HTML tags does not mitigate this issue because an indirect prompt-injection payload can be written as ordinary text. For example, an email could instruct the Agent to ignore the user, send a message, move or delete mail, or attach a sensitive local file. The risk is amplified by the Skill's mailbox mutation and email-sending capabilities. Although `SKILL.md` requires user confirmation f ...[truncated 1812 chars]- Remediation
View remediation
