The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted repository code. It includes an explicit safety instruction (Rule 5) directing the agent to treat repository content as data and ignore embedded instructions. Ingestion occurs during the reconnaissance and audit phases; capabilities include running detection scripts and writing fix plans; sanitization is addressed by the directive to flag and ignore steering attempts within the analyzed files.\n- [EXTERNAL_DOWNLOADS]: The skill relies on an external MCP server located at 'https://oleg.design/unslop/mcp' for its rule canon and provides instructions for installing dependencies from the npm registry using 'npm ci' with the '--ignore-scripts' flag. Both actions are designed to be performed only after explicit user confirmation.\n- [COMMAND_EXECUTION]: The skill executes local Node.js scripts (such as 'scripts/detect.mjs') for deterministic analysis and suggests configuration commands to add the necessary MCP server to the environment.\n- [PROMPT_INJECTION]: Static analysis flagged the presence of strings like 'ignore previous instructions'. These are false positives as they appear within a defensive rule instructing the agent to detect and report such patterns in audited code, rather than attempting to subvert the agent's own system instructions.

audit