The Agent Skills Directory

[DATA_EXFILTRATION]: The skill captures comprehensive logs of all session activity, including tool names, inputs, and outputs, which are stored in observations.jsonl files. This sensitive interaction history is periodically sent to the Claude Haiku model for analysis. Although it uses regex to redact secrets, the depth of interaction logging presents a data exposure risk.
[EXTERNAL_DOWNLOADS]: The instinct-cli.py script includes an import command that fetches behavior-defining 'instinct' files from user-specified HTTP/HTTPS URLs. These files contain markdown and YAML instructions that directly influence the agent's logic and decision-making processes.
[PROMPT_INJECTION]: The background observer process (observer-loop.sh) is vulnerable to indirect prompt injection because it interpolates raw interaction logs into an analysis prompt without boundary markers or instruction sanitization. An attacker could potentially inject malicious instructions via a tool output (e.g., from a website read by the agent) to influence the creation of 'instinct' files.
[COMMAND_EXECUTION]: The skill makes extensive use of shell scripts and the Python subprocess module to manage background workers, detect git repository metadata, and execute the claude CLI. These operations allow the skill to interact with the local operating system and environment extensively.

continuous-learning-v2