opencodex-gateway

HIGH W007: Insecure credential handling detected in skill instructions.

• Insecure credential handling detected (high risk: 0.90). The prompt includes explicit examples of API keys placed verbatim in config files and curl requests (Authorization: Bearer , api_key fields in JSON and toml), which would require the LLM to insert secret values directly into generated commands/configs, creating an exfiltration risk.

CRITICAL E006: Malicious code pattern detected in skill scripts.

• Malicious code pattern detected (high risk: 0.90). The project contains high-risk features that can be abused for deliberate malicious activity: automatic patching of user config, forwarding/storing API keys, uploading screenshots to configurable third‑party vision endpoints (data exfiltration risk), SSE logs that may leak sensitive content, and a Computer Use extension that demonstrates running arbitrary shell/AppleScript (remote code execution/backdoor potential) if requests or dashboard inputs are not strictly validated.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.90). OpenCodex proxies to configurable third‑party LLM endpoints at runtime (e.g., https://api.openai.com/v1, https://api.deepseek.com, https://api.anthropic.com/v1) and injects the returned model text (e.g., Vision Bridge descriptions and completions) directly into prompts, so remote content controls agent behavior.