The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses a bash helper script (agent-browser-sessionctl) to manage profile directories and execute the agent-browser CLI. The script employs strict input validation (regex and case matching) to prevent command injection or path traversal attacks.\n- [DATA_EXPOSURE]: The skill manages browser data in a dedicated root directory (~/.agents/chrome-profiles). It explicitly forbids reading sensitive browser stores such as cookies, login data, or history unless specifically requested for redacted diagnostics. It also enforces restrictive directory permissions using umask 077 and chmod 700.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes external arguments passed to the run command, which represents an attack surface for indirect instructions. However, the script implements a validation layer (validate_run_args) to prevent the agent from overriding security-critical flags like --session or --profile. Mandatory Evidence: (1) Ingestion point in AGENT_BROWSER_ARGS via cmd_run, (2) Boundary markers defined in SKILL.md regarding profile isolation, (3) Capability inventory includes file system operations (rm, mkdir) and CLI execution, (4) Sanitization performed via strict filename validation and flag filtering.

vp-agent-browser-session