The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The script scripts/show_snapshot.py uses pickle.loads() to deserialize evolution results from snapshot files. Deserializing untrusted pickle data can lead to arbitrary code execution. The implementation includes a mandatory CLI flag (--i-trust-this-file) and explicit warnings to mitigate this risk, but the underlying mechanism remains inherently unsafe.
[EXTERNAL_DOWNLOADS]: The skill downloads the evolutionary search engine from the official repository of Imbue, a well-known AI research organization. This behavior is documented and central to the skill's purpose.
[COMMAND_EXECUTION]: The skill frequently invokes shell commands via uv run and git to manage the lifecycle of the evolution experiments and their dependencies.
[PROMPT_INJECTION]: The skill exhibits a surface for indirect prompt injection by design, as it involves an LLM-driven loop that generates and subsequently executes or evaluates code and prompts.
Ingestion points: LLM-generated artifacts in scripts/parrot_openrouter.py and templates/custom_problem_template.py during the mutation phase.
Boundary markers: Absent; the templates do not utilize specific delimiters to isolate LLM-generated instructions from the evaluation environment.
Capability inventory: The skill enables the creation and execution of custom organisms which may perform network operations, file system access, or code execution depending on the user's problem definition.
Sanitization: No automated sanitization or verification of the LLM-generated code is performed before execution.

darwinian-evolver