The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructions direct the agent to resolve an 'active paper' and explicitly follow instructions found in files like AGENTS.md or scientific notes. This introduces a surface where malicious instructions could be embedded in these external files to influence the agent's behavior.
Ingestion points: SKILL.md (Procedure for resolving the active paper and reading its scientific notes or AGENTS.md before editing files).
Boundary markers: The instructions do not specify any delimiters or safety guidelines to help the agent distinguish between data and potentially malicious instructions in these external sources.
Capability inventory: Although the skill itself contains no executable code, it guides an agent that is expected to read/write files and potentially interact with local simulators or remote quantum hardware backends.
Sanitization: No sanitization, validation, or filtering of the content from these external project files is mentioned.

rk-quantum