The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill provides explicit instructions to install system-level dependencies using apt-get install in examples/electron.md and examples/playwright.md. This requires elevated privileges and permanently modifies the execution environment.
[COMMAND_EXECUTION]: The skill relies on complex shell command patterns to manage the application lifecycle, including backgrounding processes with &, polling network ports with curl, and managing interactive terminal sessions using tmux send-keys and tmux capture-pane.
[DYNAMIC_EXECUTION]: The Electron driver skeleton in examples/electron.md utilizes dynamic execution patterns:
It uses node -e to execute inline JavaScript for launch verification.
The REPL driver uses page.evaluate(expr) to execute arbitrary JavaScript strings—supplied by the agent—within the application's browser context.
It builds shell commands via string concatenation for execution in tmux sessions.
[INDIRECT_PROMPT_INJECTION]: The skill creates a feedback loop where the agent is instructed to read and act upon the output of the application it is driving, which represents a vulnerability surface for instructions embedded in application data.
Ingestion points: Application logs (/tmp/api.log), terminal screen content (tmux capture-pane), and web page content/DOM (page.evaluate, chromium-cli).
Boundary markers: None are defined to separate application output from the agent's internal instructions.
Capability inventory: The agent has extensive capabilities including file writing, shell command execution, and system-level package installation.
Sanitization: There is no evidence of sanitization, filtering, or escaping applied to the data ingested from the application before it is processed by the agent.

run