The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill creates a surface for indirect prompt injection by dispatching an auditor subagent to process external artifacts.
Ingestion points: The subagent reads content from {PROOF_RECORD_PATH} and {ARTIFACT_PATH} as defined in references/auditor-prompt.md.
Boundary markers: There are no explicit delimiters or instructions provided to the subagent to ignore potentially malicious instructions embedded within the code artifacts or proof records.
Capability inventory: The auditor subagent is granted the ability to read files, perform audits, and execute verification commands.
Sanitization: There is no evidence of sanitization or filtering of the content within the artifacts before they are ingested by the subagent context.
[DYNAMIC_EXECUTION]: The skill facilitates the execution of arbitrary commands generated during the build and attack phases.
In references/auditor-prompt.md, the subagent is instructed to "Execute the verification pointers in the verdict table where feasible (run the commands...)".
These commands are dynamically generated based on the agent's interaction with the artifact and stored in the proof record for later execution by the auditor.

prove-it