autobrowse

CRITICAL E006: Malicious code pattern detected in skill scripts.

• Malicious code pattern detected (high risk: 1.00). The code intentionally packages local trace artifacts (which can include cookies, auth headers, bearer tokens, private URLs) into LLM prompts and sends them to Anthropic, and it executes LLM-generated scripts and runs npm installs with the host environment — creating clear data-exfiltration and remote-execution / supply‑chain risk.

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 0.75). The required runtime workflow reads outsider-authored free text from the target website/browser (e.g., CDP Console.messageAdded / Runtime.exceptionThrown text) into unified-events.jsonl, which the outer agent then ingests into LLM context during trace reading and hypothesis/strategy updates.