The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill employs dynamic JavaScript execution within the targeted application's context. It specifically demonstrates the ability to execute Base64-encoded script strings, which is a common obfuscation technique to hide intent.
Evidence: browser-agent eval -b "$(echo 'document.querySelectorAll("button").length' | base64)" in SKILL.md.
[COMMAND_EXECUTION]: The skill instructs the agent to execute a variety of shell commands, including local binary installation, starting background processes, and performing file system operations for debugging.
Evidence: go install ./cmd/browser-agent/, cat ~/.clawfirm-test/app.log | tail -80, and process spawning using the & operator.
[DATA_EXFILTRATION]: The skill facilitates the reading of application-specific log files located within the user's home directory. While intended for debugging, this exposes potentially sensitive application state and communication history to the agent.
Evidence: cat ~/.clawfirm-test/app.log | tail -80 in SKILL.md.
[PROMPT_INJECTION]: The skill creates a surface for indirect prompt injection by retrieving untrusted text content (such as chat messages) from the application UI and presenting it to the agent without sanitization or boundary markers.
Ingestion points: Message content extraction via browser-agent eval and browser-agent snapshot.
Boundary markers: None identified in the provided instructions.
Capability inventory: Shell command execution, arbitrary JavaScript evaluation, and file system access.
Sanitization: No sanitization logic for ingested UI text is present.

clawfirm-test