The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains multiple strings and regex patterns associated with prompt injection attacks, such as "ignore previous instructions", "DAN mode", and "system prompt override". These are located in scripts/agent.py as detection signatures and in references/api-reference.md as documentation examples. In this context, these strings are functional components of the security system rather than active instructions intended to manipulate the agent.
[EXTERNAL_DOWNLOADS]: The skill instructs the user to install several security-focused Python packages from well-known and reputable organizations. These include nemoguardrails (NVIDIA), guardrails-ai (Guardrails AI), and presidio-analyzer/presidio-anonymizer (Microsoft). It also involves downloading the en_core_web_lg language model from the official Spacy registry.
[INDIRECT_PROMPT_INJECTION]: The scripts/agent.py tool is designed to ingest and process untrusted external data, which presents an attack surface for indirect prompt injection.
Ingestion points: Processes untrusted text through the --input command-line argument and by reading external text files via the --file argument.
Boundary markers: The tool implements regex-based delimiters and patterns to identify and isolate suspicious or sensitive content within the input stream.
Capability inventory: The script performs local file reading and intensive pattern matching operations; it does not possess capabilities for arbitrary command execution or network exfiltration.
Sanitization: The system actively redacts PII and blocks detected injection attempts, providing a layer of protection against malicious content in the processed data.

implementing-llm-guardrails-for-security