The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external artifacts, such as source code and design documents, to identify vulnerabilities. This process exposes the agent to potential indirect prompt injection attacks where malicious instructions in the input files could manipulate the review outcome.
Ingestion points: Artifacts are read in the initial classification phase and throughout the various attack waves.
Boundary markers: The skill employs isolated sub-agent calls (using the Agent tool) to separate challenger processes from the main session context.
Capability inventory: The skill has access to Bash, Agent, Read, Write, and WebSearch tools, which could be exploited if an injection succeeds.
Sanitization: Organizational constraints are sanitized before being passed to external sidecar models to prevent data leakage.
[COMMAND_EXECUTION]: The skill uses shell commands to invoke various local scripts and external CLI tools for multi-team analysis.
Evidence: Scripts in SKILL_detail.md show invocations of gemini, ollama, agy, and gh copilot via bash. It also runs local utility scripts like target_freeze.sh and sidecar_wait.sh.
[EXTERNAL_DOWNLOADS]: The skill utilizes npx to execute packages from the vendor and other well-known services.
Evidence: The tool invokes @chrono-meta/fh-gate and @openai/codex using npx. These are associated with the skill's vendor or established technology providers and are considered part of the expected operational environment.

steel-quench