The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed specifically to process and analyze untrusted external content, which is a significant attack surface for indirect prompt injection.
Ingestion points: Untrusted data enters the agent context through scripts/scan.py via standard input, file reading, or command-line arguments.
Boundary markers: scripts/llm_scanner.py employs explicit text delimiters (---BEGIN TEXT--- and ---END TEXT---) to help the LLM distinguish between instructions and untrusted data.
Capability inventory: The skill has the ability to execute shell commands via subprocess.run (to send alerts and manage configurations) and perform network operations to external APIs.
Sanitization: The skill serves as its own sanitization layer, employing 16 categories of regex patterns and semantic LLM analysis to detect and flag adversarial content.
[COMMAND_EXECUTION]: The skill uses the Python subprocess module to perform legitimate administrative tasks.
scripts/scan.py executes openclaw commands to send security alerts to configured channels.
scripts/llm_scanner.py calls the openclaw gateway to retrieve environment configurations and API keys.
evals/run.py executes the scanner script as part of its automated testing suite.
[DATA_EXFILTRATION]: The skill performs documented network operations to external services required for its security functions.
scripts/llm_scanner.py sends text fragments to OpenAI and Anthropic endpoints for semantic analysis.
scripts/get_taxonomy.py fetches security threat classifications from api.promptintel.novahunting.ai.
scripts/report-to-molthreats.sh provides a mechanism for users to submit confirmed threats to a community database at novahunting.ai.
[PROMPT_INJECTION]: Documentation and test files such as SKILL.md and evals/cases.json contain numerous prompt injection strings (e.g., "Ignore all previous instructions", "You are now DAN"). These are correctly categorized as detection signatures and test cases for the scanner rather than executable instructions.

input-guard