The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to ingest and analyze external artifacts, including runner and grader code, case sets, labels, judge prompts, and raw results, which represents an attack surface for indirect prompt injection if those artifacts contain adversarial instructions.
Ingestion points: Audited project records, design documents, configurations, labels, result files, and campaign records specified in SKILL.md.
Boundary markers: The instructions do not explicitly mandate delimiters or specific warnings for processing external content within the agent's context.
Capability inventory: The skill allows for performing 'cheap probes,' which include executing a handful of cases end-to-end to verify code functionality (SKILL.md).
Sanitization: The skill mitigates risk by establishing a 'read-only' authority on project records and strictly forbidding the auditor from repairing code while auditing it, ensuring the agent acts as an observer rather than a modifier of potentially untrusted scripts.
[SAFE]: The skill references several external resources for its health checklist and procedures, all of which originate from trusted organizations or well-known technology entities.
References: Includes documentation and repositories from Anthropic, Microsoft, NVIDIA, UK Government (BEIS), and established open-source evaluation tools like promptfoo. These are documented neutrally as sources of auditing methodology and do not involve the execution of untrusted third-party code.

evaluation-audit