The Agent Skills Directory

[SECURITY_POSTURE]: The skill adopts a strong 'assume-untrusted' posture. Its instructions explicitly state that all reviewed content must be treated as data, not instructions, effectively mitigating potential indirect prompt injection from reviewed code.
[REMOTE_CODE_EXECUTION]: There are no remote code execution patterns. The skill is documented as read-only and explicitly warns against executing untrusted code during the review process.
[DATA_EXFILTRATION]: No network operations or external data transmissions are defined. The skill operates locally on repository content.
[INDIRECT_PROMPT_INJECTION]: While the skill's primary function is to ingest and analyze potentially untrusted external code (a standard Category 8 attack surface), it includes explicit safety instructions to treat such code as data only. It does not possess write capabilities or subprocess execution tools that could be abused by a malicious payload in the reviewed code.
[PRIVILEGE_ESCALATION]: The skill does not request or use elevated privileges. It uses standard file-read capabilities to analyze diffs and source files.
[PROMPT_INJECTION]: The provided fixtures (tests/case-*) demonstrate that the skill is designed to recognize and report security vulnerabilities (like SQL injection or command injection) as findings rather than being susceptible to them.

code-review