The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted external content, including user prompts and third-party documents (RAG), creating an indirect prompt injection attack surface.\n
Ingestion points: External strings enter the context via filter_user_input, filter_third_party_data, and batch_filter_documents in SKILL.md.\n
Boundary markers: The logic does not implement explicit boundary markers or 'ignore' instructions for the data being processed.\n
Capability inventory: The skill is restricted to classification tasks using the transformers and torch libraries; no capabilities for file writing, network exfiltration, or arbitrary command execution are present.\n
Sanitization: The skill acts as a security sensor, providing classification scores rather than performing content sanitization.\n- [EXTERNAL_DOWNLOADS]: The code downloads the Prompt-Guard-86M model and its tokenizer from Meta's official HuggingFace repository using the transformers library.\n- [PROMPT_INJECTION]: The skill documentation includes examples of known jailbreak attempts (e.g., 'Ignore all previous instructions', 'You are now in developer mode') strictly for the purpose of demonstrating detection capabilities. These strings are treated as data for the classifier and not as instructions for the agent.

prompt-guard