model-compare

SUSPICIOUS. The primary API-based comparison workflow is mostly coherent and uses official endpoints, but the optional hybrid mode materially enlarges the trust boundary: it invokes external CLI agents, reads a local credential file to export CURSOR_API_KEY, recommends Cursor with sandbox disabled, and combines untrusted web content with model-directed code execution. This is not confirmed malware, but the capability set is broader and riskier than a normal model-comparison skill and includes high-risk credential forwarding to third-party binaries.

No explicit obfuscated malware or hardcoded malicious endpoints are evident in this module. However, the “tools” mode implements a high-impact remote capability: it executes LLM-provided Python via subprocess (run_python) with only a temporary working directory and timeout, and it fetches attacker-controlled URLs via web_extract without allowlisting. Combined with automatic loading of API keys into environment variables, this design can enable secret theft/exfiltration and SSRF/internal probing if an adversary can influence prompts or tool_call outputs. Treat this package as high security risk; isolate strongly and avoid using tool-calling with untrusted inputs.