molted
Pass
Audited by Gen Agent Trust Hub on Nov 13, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from incoming emails, which provides a surface for indirect prompt injection attacks.
- Ingestion points: The agent ingests data through tools like
read_inboxandget_thread, as well as REST endpoints such asGET /v1/agent/threads/:idandGET /v1/agent/thread-contextwhich return email bodies and subjects. - Boundary markers: The Molted infrastructure provides structural markers including
contentTrust(e.g.,"untrusted"for sender content) andinjectionRisklevels (low, medium, high) to help the agent distinguish between instructions and data. It also implements a Canary Token system (MLTED-<hex>) in the agent context to detect and block exfiltration attempts in outbound replies. - Capability inventory: The skill possesses extensive write capabilities, including sending new emails (
send_email), replying to threads (reply), and scheduling follow-ups (schedule_followup), which could be abused if an injection succeeds. - Sanitization: The platform includes an automated safety classifier that screens inbound mail for prompt injection, phishing, and malware, providing a
safetyVerdictand performing automatic quarantining of high-risk messages before they reach the agent.
Audit Metadata