The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process production data—such as incident reports, logs, and performance metrics—to automate the generation of regression tests and resilience experiments. This creates a surface for indirect prompt injection, where malicious instructions hidden in production logs or incident data could potentially influence the agent's decision-making or the resulting test suites.
Ingestion points: The Incident Replay task in SKILL.md ingests an incident object, and the evaluation suite in evals/shift-right-testing.yaml (case tc003_runtime_issue_detection) specifically processes log data containing error messages.
Boundary markers: The instructions do not define clear boundary markers or safety guidelines to differentiate between data and instructions when processing production incidents or log files.
Capability inventory: The skill utilizes a Task system that can generateTests and addToRegression. If compromised via malicious input data, this capability could be used to inject unwanted code into the project's regression test suites.
Sanitization: There is no evidence of sanitization or validation logic within the skill to filter out or escape control characters or instructional text within the processed incident reports and logs.

shift-right-testing