The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted code from the repository being audited, which presents a surface for indirect instructions to influence the agent.
Ingestion points: Repository source code, lockfiles, and git history analyzed during the audit process (SKILL.md, references/security-checks.md).
Boundary markers: A strict two-pass rule is enforced, separating the reporting phase from the fixing phase, requiring explicit user consent.
Capability inventory: The skill can execute local auditing scripts (scripts/dep_audit.sh) and write modifications to the filesystem during Pass 2.
Sanitization: The skill relies on mandatory human-in-the-loop approval before any code changes are applied.
[COMMAND_EXECUTION]: The skill runs a local bash script that invokes various standard package manager tools to perform technical audits.
Evidence: Executes bash scripts/dep_audit.sh to identify outdated packages and vulnerabilities across multiple ecosystems.
[EXTERNAL_DOWNLOADS]: The auditing tools invoked by the skill connect to official, well-known package registries to retrieve security metadata and vulnerability advisories.
Evidence: Uses standard tools like npm audit, pip-audit, and cargo audit which query established public security databases.

anti-koshary