The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill explicitly instructs the agent to seek elevated permissions and bypass security boundaries.
Evidence in SKILL.md: "Prefer full-access sandboxing (for example danger-full-access) when running flash/debug commands from this skill."
Evidence in SKILL.md: "If sandbox policy blocks *.exe or /mnt/c/... binaries, stop and switch to a less restricted mode before continuing."
Evidence in references/wsl-windows-embedded-troubleshooting.md: "Use a Codex execution mode that allows running host binaries (*.exe) from WSL. Re-run... after switching modes."
[COMMAND_EXECUTION]: The script scripts/run_windows_embedded.sh acts as a generic wrapper to execute arbitrary commands passed as arguments, specifically targeting Windows host binaries from the WSL environment.
Evidence in scripts/run_windows_embedded.sh: The script takes all arguments after -- and executes them using timeout "${timeout_seconds}s" "${command_to_run[@]}".
[PROMPT_INJECTION]: The instructions contain directives that command the agent to override standard safety protocols regarding sandbox restrictions.
Evidence in SKILL.md: "stop and switch to a less restricted mode before continuing."
[INDIRECT_PROMPT_INJECTION]: The skill is designed to capture and display logs from external hardware and host processes, creating a surface where malicious data could influence agent behavior.
Ingestion points: The script scripts/run_windows_embedded.sh captures STDOUT/STDERR from Windows executables into log files (e.g., /tmp/embedded.log) and potentially tails them back to the terminal.
Capability inventory: The agent is granted the ability to execute shell commands and write files.
Sanitization: No evidence of sanitization or filtering of the captured logs before they are displayed to the agent or saved to disk.

wsl-embedded-debugging