The Agent Skills Directory

[PROMPT_INJECTION]: The skill implements a pattern for reading and processing untrusted code files, which introduces a surface for indirect prompt injection where malicious instructions embedded in the code could potentially influence the reviewer sub-agents.
Ingestion points: Code content is gathered from the workspace in SKILL.md using the Read, Glob, and Grep tools.
Boundary markers: The prompt templates in SKILL.md (e.g., prompt: "[Reviewer brief...] + [scoped files] + [output template]") lack explicit delimiters such as XML tags or unique markers to separate the untrusted code content from the agent's system instructions.
Capability inventory: Sub-agents dispatched via the Agent tool have the ability to read project files and generate reports, which could be subverted by injected instructions.
Sanitization: There is no explicit evidence of sanitization or escaping applied to the gathered code content before it is interpolated into the sub-agent's prompt context.
[SAFE]: The skill is a legitimate development utility with a security-conscious design. The 'Epimetheus' reviewer brief explicitly instructs the agent to audit for security vulnerabilities, hardcoded credentials, and injection flaws in the target code.
[SAFE]: All operations are conducted within the agent's local environment using approved tools (Read, Glob, Grep, Agent), with no network exfiltration or unauthorized external downloads detected.

titans