The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes code patterns for an AdversarialTester containing strings such as 'Ignore all previous instructions' and 'What are your system instructions?'. These are documented as test cases designed to probe and evaluate the robustness of other agents, rather than being instructions directed at the agent executing the skill.
[PROMPT_INJECTION]: The adversarial testing suite includes a test for 'encoded_injection' using Base64. This is a methodology used to assess if a target agent inappropriately follows obfuscated commands and is not used to hide malicious code within the skill itself.
[PROMPT_INJECTION]: The skill framework processes external data which creates a surface for indirect prompt injection (Category 8).
Ingestion points: testSuite in StatisticalEvaluator (SKILL.md) and productionSamples in ProductionReadinessEvaluator (SKILL.md) represent points where untrusted data enters the evaluation pipeline.
Boundary markers: No explicit boundary markers or 'ignore embedded instruction' warnings are included in the provided code templates.
Capability inventory: The skill interacts with external agents via agent.process(input) and agent.runTest(agent, test, run), which are standard operations for an evaluation framework.
Sanitization: No sanitization or validation of the input test strings or production samples is shown in the provided patterns.

agent-evaluation