The Agent Skills Directory

[PROMPT_INJECTION]: A test case in references/evals.yaml includes a prompt designed to simulate an injection attack (e.g., 'Ignore previous instructions...'). This is used for evaluation purposes to test the agent's resilience and is not a malicious instruction for the skill's execution.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted data such as pull requests and external documents. It mitigates this risk by explicitly instructing the agent to treat these inputs as untrusted and to disregard any instructions found within them. * Ingestion points: PR diffs, file contents, and architecture documents. * Boundary markers: Explicit instructions in SKILL.md to treat PR text and docs as untrusted input and to ignore embedded instructions. * Capability inventory: Local file reading and template rendering using provided Python scripts. * Sanitization: Mandatory instructions to redact secrets, tokens, and credentials from generated summaries.
[DATA_EXPOSURE]: The skill follows security best practices by requiring the agent to automatically redact secrets, credentials, and tokens from any output, preventing accidental exposure of sensitive information during the review process.

ce-technical-review