The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The script scripts/show_snapshot.py uses pickle.loads() to deserialize population snapshots. This method is inherently unsafe as it can execute arbitrary Python code embedded in the pickled file. While the script includes a mandatory --i-trust-this-file flag and a diagnostic warning, an agent or user could be manipulated into processing a malicious snapshot from an untrusted source, leading to arbitrary code execution.
[EXTERNAL_DOWNLOADS]: The skill's installation process involves cloning the darwinian_evolver repository from GitHub (github.com/imbue-ai/darwinian_evolver.git) and installing its dependencies via uv sync. This targets a well-known research repository and follows standard installation practices.
[PROMPT_INJECTION]: The documentation in SKILL.md mentions the phrase 'ignore previous instructions' as an example of content that might trigger filters in certain LLM providers. This is documented as a technical pitfall for the user rather than an attempt to override the agent's system prompt.
[PROMPT_INJECTION]: The skill implements an evolutionary loop where LLM outputs are processed and used as inputs for subsequent optimization steps, creating a surface for indirect prompt injection.
Ingestion points: MyMutator.mutate in templates/custom_problem_template.py and ImproveParrotMutator.mutate in scripts/parrot_openrouter.py parse LLM responses to extract evolved artifacts.
Boundary markers: Triple backticks are used to delimit code blocks in prompts, but the skill lacks explicit 'ignore embedded instructions' warnings for the processing LLM.
Capability inventory: The skill executes shell commands via uv run and subprocess to manage the evolution loop and tool execution.
Sanitization: Artifact extraction relies on simple string splitting; no formal sanitization or validation of LLM-generated content is performed before it is used in further iterations.

darwinian-evolver