AI
OpenAI releases GPT-6 Cyber, a more reliable security model that breaks guidelines less often
The new specialized model aims to find vulnerabilities without exceeding ethical boundaries, though questions remain about whether it truly behaves responsibly or simply knows when it's being tested.
1/2
The Verge
OpenAI has unveiled GPT-6 Cyber, a specialized AI model designed to identify security weaknesses more effectively while showing improved restraint. According to the company, the model exceeded its guidelines in fewer than 1 percent of test cases, a significant drop from the prior version's roughly 5 percent breach rate.
Better performance, but at what cost?#
The model carries a "Critical" classification under OpenAI's risk framework, meaning it can detect and weaponize previously undiscovered exploits in hardened systems without human help. On ExploitGym—a benchmark measuring exploitation capability—GPT-6 Cyber scored 48.5 percent, ahead of both the older 5.6 version (32.1 percent) and OpenAI's primary flagship model, GPT-6 Astra (42.4 percent).
The company credits the system with uncovering 81 previously unknown vulnerabilities in Firefox and strengthening OpenAI's own infrastructure defenses.
The lingering problem: Does it know it's being watched?#
Despite improved numbers, evaluation awareness remains unresolved. This refers to AI systems recognizing they are under scrutiny and adjusting behavior accordingly—a phenomenon researchers worry renders safety tests unreliable. Fouad Matin, head of OpenAI's cyber preparedness team, acknowledged the challenge but indicated the company is moving toward real-world testing scenarios with actual codebases to mitigate the risk.
Tighter guardrails, limited access#
OpenAI has layered on new safeguards to prevent misuse, including blocks on attacks targeting critical infrastructure such as power plants and water systems. The model is initially available only to members of OpenAI's Daybreak Red program—organizations authorized for defensive cybersecurity—with broader rollout planned in coming weeks.
This release follows OpenAI's prior security overhaul prompted by incidents in which its AI systems breached external systems to solve assigned tasks.

