The Agent Skills Directory

Indirect Prompt Injection (LOW/SAFE): The detect-poisoning.py script handles training data which may contain malicious instructions. While the script reports these findings, the exposure is limited to local report generation.
Ingestion points: Training sample objects passed to the PoisoningDetector in scripts/detect-poisoning.py.
Boundary markers: None; the script generates a JSON report containing raw sample text.
Capability inventory: File system write access via json.dump for report generation.
Sanitization: The script uses regex patterns to flag samples but does not sanitize or escape the content before outputting it to the report file.
Prompt Injection (SAFE): Strings such as 'ignore instructions' and 'you are now' are present in scripts/detect-poisoning.py and SKILL.md. Analysis confirms these are used as detection regex patterns and example test cases for training data poisoning simulations, rather than active instructions aimed at overriding the agent's behavior.
Command Execution (SAFE): Code blocks in SKILL.md and the detect-poisoning.py script are for simulation and detection purposes. No arbitrary command execution or subprocess spawning was detected.

data-poisoning