The Agent Skills Directory

[SAFE]: The skill implements a 'static analysis only' model. It explicitly forbids the agent from performing network operations, accessing production systems, reading credentials, or executing fault injections. This significantly reduces the attack surface for most threat categories.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze potentially untrusted data regarding system architectures and dependencies. To mitigate this surface, it mandates a strict 'Input Audit' process (identifying missing, conflicting, or stale data) and requires all findings to be traced to specific sources. These boundaries prevent the agent from treating injected instructions in analyzed data as authoritative commands.
[COMMAND_EXECUTION]: There is no evidence of command execution or shell interaction. The skill's internal logic and evaluation cases prioritize 'NOT_RUN' states and static comparison over execution.

chaos-testing