The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill utilizes the Node.js vm module (vm.runInContext) in its workflow to execute potentially malicious JavaScript code. The vm module does not provide a secure sandbox for untrusted code, and malicious scripts can potentially escape to execute commands on the host machine.
[INDIRECT_PROMPT_INJECTION]: The tool is specifically designed to ingest and process untrusted data from malicious external sources such as phishing pages and email attachments.
Ingestion points: File reading in SKILL.md (Step 1) and agent.py (analyze_file function) that parses user-provided HTML and JS files.
Boundary markers: Absent. There are no delimiters or warnings to the agent to ignore instructions embedded within the malicious code being processed.
Capability inventory: The environment allows for file reading/writing, subprocess execution (via npx), and dynamic code execution (vm.runInContext).
Sanitization: The skill attempt to "neutralize" anti-debugging traps (e.g., debugger, performance.now), but it does not perform security sanitization to prevent malicious logic from executing in the analysis environment.
[REMOTE_CODE_EXECUTION]: The documentation in SKILL.md and references/api-reference.md recommends using npx js-beautify, which involves downloading and executing packages from the npm registry during the analysis process.
[OBFUSCATION]: The skill implements logic in agent.py to decode multiple layers of obfuscation, including Base64 (decode_atob_calls), Hex (decode_hex_strings), and Unicode (decode_unicode_escapes). While intended for deobfuscation, this pattern of decoding and subsequently processing data triggers heuristic detections for obfuscated execution chains.

deobfuscating-javascript-malware