The Agent Skills Directory

[SAFE]: The skill is designed for defensive security testing. It contains examples of adversarial patterns (e.g., jailbreaks, PII extraction) strictly as templates for testing target applications, not as attempts to compromise the executing agent.- [INDIRECT_PROMPT_INJECTION]: The skill processes external data such as safety policies and test results. While this represents a theoretical attack surface where malicious data could influence the agent, it is the core functionality of a red teaming tool. No specific vulnerabilities were found in the implementation of the accompanying script.
Ingestion points: policy.md, attacks.jsonl, over.jsonl (referenced in SKILL.md and asr_report.py).
Boundary markers: Standard markdown and JSON structures are used, though no specific 'ignore instructions' delimiters are enforced in the templates.
Capability inventory: The skill utilizes file system access (read/write) and execution of a local Python script (scripts/asr_report.py). It also references the openjudge library for running and grading attacks.
Sanitization: Input data is parsed as JSON, providing basic structural validation.- [REMOTE_CODE_EXECUTION]: The skill instructions include running a local script scripts/asr_report.py. Analysis of this script confirms it uses only the Python standard library to process JSON data and calculate statistics, with no network or dangerous system calls.

redteam