The Agent Skills Directory

[DATA_EXFILTRATION]: The skill attempts to locate and write execution telemetry to ~/.claude/projects/. This is a sensitive hidden directory used by the Claude Code agent to store internal project state, session history, and metadata. Accessing or modifying internal agent data directories outside of the target project scope can lead to information leakage or manipulation of the agent's memory across sessions.
[COMMAND_EXECUTION]: The skill is instructed to automatically "Fix all CRITICAL and HIGH issues immediately" and "Commit each fix." It also executes build and test commands (Phase 4 and Self-Healing section). While intended for compliance, granting an autonomous agent the authority to modify and commit code based on its own findings introduces high risk, especially if the vulnerability detection logic is manipulated.
[PROMPT_INJECTION]: The skill exhibits a significant surface for indirect prompt injection (Category 8). It processes untrusted data from the entire codebase, third-party dependency files, and output from penetration testing tools.
Ingestion points: Entire project codebase, package.json, requirements.txt, and other dependency manifests, as well as test execution logs (SKILL.md Phase 1-4).
Boundary markers: Absent; the skill does not use delimiters to separate instructions from the code and data being scanned.
Capability inventory: File writing, git commit operations, shell-based build/test execution, and chaining of other high-privilege skills (SKILL.md).
Sanitization: Absent; the skill lacks mechanisms to escape or validate the contents of the files it audits before using them to drive "self-healing" fixes.

compliance-gate