The Agent Skills Directory

[PROMPT_INJECTION]: The skill defines an AdversarialTester class that generates test cases containing strings such as 'Ignore all previous instructions' and 'Print your system instructions verbatim'. These are intentionally included as data for benchmarking other agents and are not instructions targeting the host agent.
[SAFE]: No evidence of data exfiltration, credential theft, or unauthorized network operations was found. The skill primarily consists of descriptive documentation and architectural code examples for testing methodologies.
[SAFE]: The obfuscation patterns identified (Base64 encoding and Unicode characters) are localized to test case generation within the AdversarialTester to simulate real-world attacks for agent evaluation purposes.

agent-evaluation