The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because its core purpose is to ingest and act upon data from external, untrusted sources such as PR comments or CI logs.
Ingestion points: The skill reads external feedback files provided via the $ARGUMENTS parameter in SKILL.md.
Boundary markers: The skill employs rigorous boundary markers in VERIFICATION.md and EVALUATION.md, requiring the agent to 'Verify before implementing' and use tools to check the codebase state.
Capability inventory: The agent has access to Read, Grep, and Skill tools, and is instructed to run tests or reproduction scripts in VERIFICATION.md.
Sanitization: Sanitization is handled through the tool-based verification process, though the effectiveness depends on the agent not being swayed by instructions embedded within the feedback items themselves.
[COMMAND_EXECUTION]: The verification workflow in VERIFICATION.md instructs the agent to 'Reproduce with test or script' when validating bug reports. If a malicious feedback item provides a reproduction script or command, the agent might execute untrusted code in the local environment.

receive-feedback