The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The shell scripts bot/imessage/test-imessage-bridge.sh and bot/imessage/send-imessage-test.sh accept a BlueBubbles password as a command-line argument, which exposes the credential in the system's process list. Additionally, references/agent-browser-login.md instructs users to manually copy sensitive session cookies from their browser to a local file, encouraging insecure handling of raw secrets.\n- [REMOTE_CODE_EXECUTION]: The skill makes extensive use of agent-browser eval to execute arbitrary JavaScript in a browser context. It also injects a custom monitoring probe (scripts/agent-gateway/probe-events.ts) that monkey-patches native browser functions like WebSocket and fetch to intercept and log raw event data.\n- [COMMAND_EXECUTION]: The skill relies on osascript to perform UI automation on macOS native applications (Discord, Slack, etc.) and executes various shell scripts that perform system-level operations like process management and video encoding.\n- [DATA_EXFILTRATION]: Capabilities for capturing window-specific screenshots and full-screen video recording are provided via capture-app-window.sh and record-app-screen.sh. These can be leveraged to capture sensitive data displayed on the screen.\n- [PROMPT_INJECTION]: The skill possesses a significant attack surface for indirect prompt injection. \n
Ingestion points: Browser snapshots (agent-browser snapshot -i), text retrieval (agent-browser get text), and screen captures (screencapture). \n
Boundary markers: Absent; the agent is not instructed to ignore malicious instructions embedded in the captured data. \n
Capability inventory: UI control via osascript, browser interaction via agent-browser, and network requests via curl and fetch. \n
Sanitization: Absent; external data is processed without escaping or filtering before potentially influencing agent behavior.

local-testing