The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Automated security scanners identified a Phishing alert for a referenced documentation URL (https://basarat.gitbook.io/typescript/). While hosted on a well-known platform, the specific content was flagged as malicious.
[DATA_EXFILTRATION]: The skill instructions and diagnostic scripts (specifically within the 'Toolkit Evolution' and 'Weak-Model Uplift' modes) read sensitive local data from the agent's environment, including ~/.claude/learning/usage.db, learning.db, and .claude/settings.json. This data contains private usage history and internal system configurations which are processed and potentially exposed to the model context.
[COMMAND_EXECUTION]: Several bundled Python scripts, such as run_eval.py and optimize_description.py, execute the 'claude' CLI via subprocess.run to perform automated tasks. This allows the skill to programmatically interact with the model and the filesystem using the current session's permissions.
[DYNAMIC_EXECUTION]: The 'Toolkit Evolution' and 'Agent Comparison' modes implement an autonomous optimization loop that proposes new code and prompt variants, evaluates them, and suggests or applies those changes. This represents a self-modifying system with the capability to execute generated code.
[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to scan and analyze other skill packages. This provides a significant surface for indirect prompt injection if the toolkit is directed to analyze or evaluate a malicious or poisoned skill package.

toolkit