autobrowse

No explicit malicious logic (e.g., hardcoded exfiltration endpoints, backdoor mechanisms, or obfuscated payloads) is visible in this fragment. However, it implements a security-critical generate-and-execute workflow: it sends local task/trace content to an external LLM, writes the model output to disk as a runnable script, scaffolds project files, and then (when VERIFY is enabled) executes a framework runner that consumes that generated script while inheriting process.env. This creates a high-risk attack surface if any upstream inputs (trace data), local templates/runners, or LLM outputs are not strictly trusted and sandboxed.

No direct evidence of embedded malware (no explicit exfiltration, shelling, or obfuscated payloads) is present in this module. The primary security concern is operational/supply-chain risk: it can run npm install in an attacker-influenced outDir/package.json context (with audit disabled) and then execute an attacker-specified script via npx tsx, while inheriting all parent environment variables into those child processes. Result pass/fail is also derived from untrusted stdout JSON, which a compromised script can forge.