The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides instructions for executing various command-line tools necessary for Java security auditing.
Evidence includes commands for the CFR decompiler (java -jar cfr.jar), package management tools (pip install semgrep, composer require psalm), and SAST scanners (semgrep, codeql, snyk). These are standard tools for the skill's intended purpose.
[EXTERNAL_DOWNLOADS]: The documentation recommends the use of external security tools and rule sets.
References are made to well-known and trusted services including GitHub (for CodeQL and Semgrep rules), Snyk, and SonarQube.
[INDIRECT_PROMPT_INJECTION]: As a code auditing utility, the skill is designed to process untrusted third-party source code, which represents a potential attack surface for indirect prompt injection (e.g., malicious comments in the code under audit).
Ingestion points: The agent ingests Java source code, .class files, and JAR/WAR artifacts specified by the user in SKILL.md and references/decompile-strategy.md.
Boundary markers: The 'Evidence Contract System' (references/evidence-contract.md) acts as a logical boundary, requiring the agent to provide specific data flow evidence (EVID_*) before confirming a vulnerability, which helps mitigate the impact of injection-based decoys in the source code.
Capability inventory: The skill uses analysis capabilities and suggests running external CLI tools for verification.
Sanitization: The methodology relies on human review and structured evidence rather than automated string sanitization of the audited code.

java-audit-pipeline