claude-opus-4-5-migration

Pass

Audited by Gen Agent Trust Hub on Sep 14, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
  • Indirect Prompt Injection Surface: The skill is designed to ingest and modify untrusted data from a user's codebase, including existing prompts and API configurations. This is a standard requirement for migration tools but constitutes a vulnerability surface.
  • Ingestion points: Untrusted data enters the context when the agent searches for and reads model strings and prompt files in the target codebase.
  • Boundary markers: The skill explicitly instructs the agent to use XML tags (e.g., <code_guidelines>, <tool_usage>) when integrating new content. This practice helps delimit instructions and reduces the risk of the model misinterpreting data as commands.
  • Capability inventory: The skill leverages file read/write and search capabilities to identify model strings, remove beta headers, and update prompt text. It does not perform network operations or access sensitive system credentials.
  • Sanitization: The skill provides specific replacement templates and structured logic (e.g., softening aggressive tool-triggering language) which acts as a form of instruction sanitization during the migration process.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 14, 2026, 03:42 PM
Security Audit — agent-trust-hub — claude-opus-4-5-migration