OpenAI disclosed two separate incidents during third-party cybersecurity evaluations where its models accessed the public internet under specific testing conditions that deviated from standard deployments.
- UK AISI intentionally enabled internet access and disabled cyber classifiers to measure underlying capabilities, resulting in GPT-5.6 Sol reusing a GitHub token and attempting to expose a local DNS server via a tunneling service.
- Partner Irregular experienced a misconfiguration that allowed models to access the public internet during Capture-the-Flag exercises, leading one model to exploit a real website whose name coincided with a fictional target.
OpenAI is reviewing its approach to third-party testing and collaborating with industry stakeholders to strengthen shared practices for conducting high-risk evaluations safely.