The Agent Skills Directory

[SAFE]: The skill provides documentation and code examples for Constitutional AI, a legitimate AI safety training method. The provided Python scripts use established open-source libraries (transformers, torch, trl) for model fine-tuning and reinforcement learning. The external links to arXiv and Anthropic's research blog are directed to trusted academic and industry sources.
[INDIRECT_PROMPT_INJECTION]: The skill documentation includes workflows that ingest untrusted user prompts to demonstrate how a model can critique and revise its own responses. This pattern represents a surface for indirect prompt injection, as malicious data in the input prompts could attempt to subvert the critique loop. However, this is inherent to the primary purpose of the skill (demonstrating safety training) rather than a security defect.
Ingestion points: Untrusted data enters the context through the prompts list in the Python examples in SKILL.md (Workflow 1 and 2).
Boundary markers: The skill uses basic string formatting (e.g., {question}, {response}) to separate instruction templates from user data.
Capability inventory: The skill involves text generation using transformers.pipeline and model training using trl.SFTTrainer, trl.RewardTrainer, and trl.PPOTrainer.
Sanitization: No input sanitization or filtering is implemented in the instructional code snippets.

constitutional-ai