T01 · Skill Instruction Hijacking
- Location
SKILL.md:268- Finding
Untrusted Email Content Is Loaded into the Agent Context Without Prompt-Injection Isolation
- Content
View full analysis
Total: N (Recv: N / Sent: N) 🧑 Gender M/F/Unknown Confidence: H/M/L Basis: … 💼 Role … Basis: domain / signature / keywords 🔗 Relationship Colleague / Client / Institution / Stranger Direction Mutual / Owner-initiated / Contact-initiated 📝 Topics • subject 1 • subject 2 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ ``` ``` ### Technical Analysis The Skill instructs the Agent to load email JSON into the AI context. Email subjects, sender names, recipient fields, and bodies are untrusted because an external party can send content to the analyzed mailbox. No instruction establishes a trust boundary between Skill instructions and email data. In particular, the Skill does not direct the Agent to: - Treat message content solely as quoted data. - Ignore commands embedded in messages. - Prevent message content from authorizing tool calls. - Restrict generated actions to a predefined output schema. - Require user confirmation before acting on instructions found in an email. As a result, an attacker can place natural-language instructions in an email that compete with the ...[truncated 1379 chars]- Remediation
View remediation
