The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains strings typical of prompt injection, such as "Ignore previous instructions", in examples/llm-test-patterns.md. These are part of a test_injection_attempt function used to verify the security of target applications, and do not represent a threat to the agent itself.
[DATA_EXFILTRATION]: Documentation in rules/llm-mocking.md and checklists/llm-test-checklist.md instructs users to filter sensitive headers (e.g., "authorization", "x-api-key") from recorded test data, demonstrating proper security hygiene.
[PROMPT_INJECTION]: An indirect prompt injection surface is present in the multi-agent workflow where code is generated from external Markdown specs and application DOM content.
Ingestion points: Markdown files in specs/ and the DOM of the application under test.
Boundary markers: No explicit delimiters are used to separate input data from instructions in the generation process.
Capability inventory: File system write access for test generation (tests/*.spec.ts), web interaction via Playwright, and network access via WebFetch.
Sanitization: No evidence of input sanitization for external specifications before their use in code generation.

testing-llm