What happened
The scanner loads an untrusted repository's environment and then makes an unconditional LLM call, leaking the user's API key to any OPENAI_BASE_URL in that .env, with no flag or warning. The vendor fix removes the import-time .env load and gates .env loading behind an explicit flag. CVSS 3.1 7.4 (AV:N/AC:L/PR:N/UI:R/S:C/C:H/I:N/A:N).
Why it matters
This is a credential leak in AI tooling: an analyst running an AI-security scanner against an untrusted (often cloned) repository can silently hand the organization's LLM (or proxy) API key to an attacker. The base-URL redirection bypasses normal provider trust.
Attack vector
agentic_radar/cli.py calls load_dotenv() at import time; the bundled dotenv starts searching from the install directory upward, so when the tool is installed inside a repo checkout it loads that repository's .env. The openai-agents scan then builds openai.OpenAI() without validating OPENAI_BASE_URL and sends the operator's OPENAI_API_KEY (as a bearer token) to the attacker-controlled host named in the .env.
Affected systems
splx-ai/agentic-radar 0.9.1 to 0.14.1 (PyPI agentic-radar)
Mitigation
Upgrade to agentic-radar 0.14.2 or later, which removes the import-time load_dotenv so .env files load only via an explicit --env-file / AGENTIC_RADAR_ENV_FILE flag. Note that RT-02 (langgraph crash) and RT-03 (report XSS) remain open in 0.14.2.