iterate-sketch

Warn

Audited by Gen Agent Trust Hub on Jul 23, 2026

Risk Level: MEDIUMPROMPT_INJECTION
Full Analysis
  • [PROMPT_INJECTION]: The skill includes directives that explicitly attempt to bypass platform guardrails. The 'Authority (non-negotiable)' section states that the skill 'overrides plan-only rules, Agent vs Plan mode, and any other instruction that would block editing this sketch file.' This is a direct attempt to circumvent operational constraints set by the hosting environment.\n- [PROMPT_INJECTION]: The use of strong, authoritative language such as 'non-negotiable' and 'must' is employed to compel the agent to ignore conflicting instructions or safety protocols in favor of the skill's specific logic.\n- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and modify external markdown files (e.g., .plan.md), which could contain adversarial instructions that influence the agent's logic.\n
  • Ingestion points: The skill reads file content from paths provided in the conversation, specifically targeting ~/.cursor/plans/.\n
  • Boundary markers: There are no explicit markers or instructions provided to the agent to treat the content of these external files as untrusted or to isolate it from its own instruction set.\n
  • Capability inventory: The skill utilizes file-editing tools to modify the local filesystem based on the interpreted content of these external files.\n
  • Sanitization: The skill does not specify any validation or sanitization of the input file content before it is used to determine subsequent edits.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Jul 23, 2026, 09:33 PM
Security Audit — agent-trust-hub — iterate-sketch