The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/validate.py implements a dynamic execution sink that runs code generated during the agent pipeline.
Evidence: In scripts/validate.py, the _exec_snippet function calls exec(snippet, ns) to test the validity of verification snippets.
Risk: The execution context ns explicitly includes the full __builtins__ module. This allows the executed snippet to access sensitive functions such as __import__ and open, which could be exploited to perform unauthorized file operations or execute system commands if the generated code is malicious.
[PROMPT_INJECTION]: The skill architecture is susceptible to indirect prompt injection via the untrusted "task description" input.
Ingestion points: The skill relies on an external "task description" provided at runtime to define the benchmark environment and verification constraints used by the Miner and Validator agents.
Boundary markers: The instructions lack strong delimiters or "ignore instructions" warnings to prevent the agent from obeying malicious commands embedded within the task description.
Capability inventory: The skill has the capability to write to the local filesystem (graph artifacts) and execute Python code via the validation script.
Sanitization: There is no evidence of sanitization or structural validation of the task description before it is interpolated into agent prompts, which could allow an attacker to influence the content of the executable verification snippets generated by the Miner agent.

skillhone-synthesis