The Agent Skills Directory

[PROMPT_INJECTION]: The skill uses clear, prescriptive instructions to define its operating boundaries, including explicit "Non-negotiable rules" that prevent the agent from modifying project files. It does not contain patterns attempting to bypass platform security or override system prompts.
[DATA_EXPOSURE]: The skill is designed to read potentially sensitive files such as environment manifests, dependency files, and source code. This is necessary for security auditing. There is no evidence of network tools (curl, wget) or commands being used to exfiltrate this data.
[INDIRECT_PROMPT_INJECTION]: As the skill ingests untrusted project data (PRDs, code, specs), it is inherently exposed to indirect prompt injection surfaces. However, it mitigates this risk by requiring findings to be evidence-backed and subject to a "Quality Gate" that checks for invented content and requires specific file/line references for every finding, reducing the likelihood of following malicious instructions embedded in project code.

security-pass