The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes various shell commands such as grep, find, which, and file to inspect the local system environment and binary metadata.- [DATA_EXPOSURE_AND_EXFILTRATION]: The skill accesses sensitive configuration files and shell profiles, including ~/.claude/settings.json, ~/.zshrc, ~/.bashrc, and ~/.bash_profile. The instructions direct the agent to run grep -n "--resume\|--continue" ~/.zshrc and similar commands to identify specific flags. While these files are sensitive, the access is limited to read-only string matching for diagnostic purposes and does not include mechanisms to send the data to external servers.- [INDIRECT_PROMPT_INJECTION]: The skill performs broad scans of project-level files (CLAUDE.md) across the filesystem. If these files are under the control of an attacker, they could contain malicious instructions designed to be executed during the audit process. 1. Ingestion points: Local CLAUDE.md files discovered via filesystem searches. 2. Boundary markers: Absent; the skill processes file content directly without delimiters or safety instructions. 3. Capability inventory: Execution of shell commands via bash for executing audit phases. 4. Sanitization: None; the contents of the scanned files are not validated or sanitized before being returned to the agent context.

check-cache-bugs