evolve
Warn
Audited by Gen Agent Trust Hub on Aug 14, 2026
Risk Level: MEDIUMEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTION
Full Analysis
- [UNVERIFIABLE_DEPENDENCIES_AND_REMOTE_CODE_EXECUTION]: The skill instructs the agent to execute
npx metaharness-darwin, which downloads and runs a package from the npm registry at runtime. This package is not from a recognized trusted vendor or the skill author's established resource patterns. - [COMMAND_EXECUTION]: The skill provides commands to run shell scripts via
npm run evolveandnpx, granting the agent the capability to execute arbitrary commands within the substrate environment. - [DYNAMIC_EXECUTION]: The primary purpose of the skill is 'Darwin Mode', where the harness evolves by mutating its own source files (planner, policies, etc.) and executing the resulting variants. This represents runtime generation and execution of code.
- [INDIRECT_PROMPT_INJECTION]: The skill defines a feedback loop where model-generated mutations are executed to 'improve' the harness. While a validation gate (
validateGeneratedCode) is mentioned in the text, the architecture creates an attack surface where untrusted model output directly influences the execution logic of the tool itself.
Audit Metadata