The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests data from external URLs and processes it to generate code modifications, creating an attack surface for indirect prompt injection.\n
Ingestion points: Content is fetched from remote servers via scripts/audit-meta.mjs and scripts/extract-jsonld.mjs using fetch and curl.\n
Boundary markers: The skill lacks explicit instructions or markers to distinguish between audit metadata and potentially malicious instructions embedded in the targeted HTML or JSON-LD.\n
Capability inventory: In 'Phase 3' (SKILL.md), the agent is granted permission to write code fixes to the local working directory based on the audit results.\n
Sanitization: There is no mechanism to sanitize the ingested data for natural language instructions that might manipulate the agent's behavior during the codebase modification phase.\n- [COMMAND_EXECUTION]: The skill relies on executing shell and Node.js scripts to perform audit tasks.\n
SKILL.md instructs the agent to run bash scripts/redirect-trace.sh and node scripts/audit-meta.mjs.\n
redirect-trace.sh invokes curl to trace network redirect chains.\n
These tools are intended for the skill's primary function but involve executing processes with external inputs.

discoverability