handoff

Pass

Audited by Gen Agent Trust Hub on Jul 21, 2026

Risk Level: SAFEPROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes handoff documents that contain natural language instructions, which can be used to influence the agent's behavior in subsequent sessions.\n
  • Ingestion points: The Consume Workflow in SKILL.md directs the agent to read and follow the 'Immediate Next Steps' from a handoff file.\n
  • Boundary markers: The skill recommends verifying the handoff against the live workspace state, but does not provide robust delimiters to prevent the agent from obeying adversarial instructions within the document.\n
  • Capability inventory: The agent's ability to execute commands and modify files makes it susceptible to executing malicious steps listed in a compromised handoff file.\n
  • Sanitization: No sanitization or filtering logic is applied to the handoff content before the agent adopts it as its new task list.\n- [DATA_EXPOSURE_AND_EXFILTRATION]: The skill is designed to collect and aggregate metadata about the current workspace environment.\n
  • The Produce Workflow in SKILL.md instructs the agent to gather information on the current branch, dirty files, recent commits, and active processes.\n
  • Although the skill contains explicit instructions to never include secrets or passwords, the systematic collection of repository and environment metadata increases the surface area for information exposure.
Audit Metadata
Risk Level
SAFE
Analyzed
Jul 21, 2026, 01:20 PM
Security Audit — agent-trust-hub — handoff