The Agent Skills Directory

[PROMPT_INJECTION]: Static analysis identified instructions in references/motion-audit/index.md that mention overriding system rules. Detailed review confirms these are defensive guidelines instructing the agent to treat repository content as inert data and explicitly ignore any prompt injection attempts found within processed files.
[DYNAMIC_EXECUTION]: The skill includes procedures for verifying the safety of code changes (in references/impact/blast-radius.md) by running isolated scripts or "scratch probes." The instructions emphasize maintaining a read-only posture toward production code and using isolated test states to prevent unintended side effects.
[INDIRECT_PROMPT_INJECTION]: The skill establishes robust boundaries for processing untrusted data by instructing the agent to separate repository content from its own instructional logic, reducing the surface area for indirect prompt injection attacks during code reviews.

atlas-review