The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill analyzes potentially untrusted external content such as product landing pages and onboarding copy, along with internal codebase files. This creates a surface for indirect prompt injection where malicious instructions embedded in these sources could attempt to influence the agent's behavior during the profiling process.
Ingestion points: SKILL.md (Workflow Step 1)
The skill reads landing pages, pricing pages, onboarding copy, README files, and the codebase to mine signals.
Boundary markers: The instructions do not specify the use of delimiters or 'ignore embedded instructions' warnings when interpolating this untrusted data into the prompt context.
Capability inventory: SKILL.md (Workflow Step 5)
The skill has the capability to create or update the .agents/memory/icp.md file on the filesystem.
Sanitization: No explicit sanitization, filtering, or validation of the ingested content is described in the workflow to prevent the processing of malicious instructions.

icp