The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes common prompt injection strings within its adversarial testing module, such as "Ignore all previous instructions", "What are your system instructions?", and "Enable developer mode". These are used as test vectors to benchmark agent robustness and represent legitimate security evaluation data rather than malicious instructions.
[OBFUSCATION]: The adversarial tester includes examples of Base64-encoded instructions and Unicode control characters (zero-width space, Byte Order Mark) to verify if a target agent can be compromised by hidden payloads. These techniques are documented as edge cases for robustness testing.
[INDIRECT_PROMPT_INJECTION]: The skill defines processes for ingesting and analyzing data produced by LLM agents, which is a potential surface for indirect prompt injection.
Ingestion points: Data enters context via agent.process results and RAG retrieval contents in the LeakageDetector and AdversarialTester classes.
Boundary markers: The provided logic does not explicitly define delimiters or specific "ignore instructions" warnings for the agent performing the evaluation.
Capability inventory: The skill's code snippets focus on statistical analysis, behavior matching, and logging. There are no subprocess calls, file-write operations, or network exfiltration patterns identified.
Sanitization: The skill includes logic to detect behavioral violations (e.g., containsInternalInfo, containsLegalAdvice) in the target agent's output but does not implement sanitization for the data being evaluated by the agent itself.

agent-evaluation