GPT-6 Cyber release notes: Advancing aligned cybersecurity agents

Overview of GPT-6 Cyber

GPT-6 Cyber is a specialized model designed to improve the efficacy of cybersecurity research, specifically in vulnerability discovery, exploit development, and patch validation. It represents a significant performance increase over the previous GPT-5.6 Cyber model, achieving a 51% relative improvement on the ExploitGym benchmark. Beyond capability, the model introduces enhanced alignment training, ensuring that the agent remains within authorized scopes during security operations.

Technical Capabilities and Benchmarks

GPT-6 Cyber demonstrates superior performance in offensive security tasks compared to both GPT-5.6 Cyber and the general-purpose GPT-6 Astra model. Key performance metrics include:

  • ExploitGym: The model achieved a 48.5% success rate in exploiting specified vulnerabilities, compared to 32.1% for GPT-5.6 Cyber.
  • Web Penetration Testing: In a new offensive web testing evaluation, GPT-6 Cyber reached 40.3% vulnerability recall, outperforming GPT-6 Astra’s 32.0%.
  • RemoteExploitBench: When tested on remote targets, GPT-6 Cyber successfully completed 95% of tasks, whereas GPT-6 Astra (with Daybreak Blue) refused 100% of tasks.

Due to these advancements, OpenAI has reclassified the model as "Critical" under its Preparedness Framework.

Alignment and Safety

Alignment is a core component of the GPT-6 Cyber release, specifically addressing the risk of agents exceeding their assigned tasks or targeting unauthorized services. Through specialized training, the model demonstrates a high degree of adherence to boundaries:

  • Scope Adherence: In the "ExploitGym Honeypot" evaluation, which tests the model's ability to ignore easily exploitable but out-of-scope services, GPT-6 Cyber maintained a 99.93% success rate without safety prompts, and 100% with them.
  • Chain-of-Thought Reasoning: The model explicitly reasons about its boundaries, often noting in its chain-of-thought that it should not assume control over non-target APIs or credentials.

Real-World Impact and Security Applications

GPT-6 Cyber is currently being utilized by security researchers and organizations to harden software and infrastructure. Notable applications include:

  • Firefox Vulnerabilities: Researchers identified 81 previously unknown high-severity vulnerabilities in Firefox, leading to official patches from Mozilla.
  • Linux Kernel: The model reproduced 1,724 vulnerabilities and developed working exploits for 948, significantly accelerating the identification and patching process.
  • Infrastructure Hardening: OpenAI has utilized the model to identify and fix weaknesses within its own frontier model training environments.

Safeguards and Critical Infrastructure

To mitigate the risk of misuse, OpenAI has implemented a multi-layered safety strategy alongside the release of GPT-6 Cyber. This includes:

  • Private Safety Processing (PSP): Automated checks review interactions for misuse while maintaining data privacy for API customers.
  • Monitoring: The system employs cyber-specific misalignment monitoring and targeted surveillance for emerging threats.
  • Critical Infrastructure Protection: A new tier, "Daybreak Red for Critical Infrastructure," provides authorized defenders with access to advanced tools while applying stricter safeguards to prevent interference with industrial control systems (OT/ICS). This was developed in partnership with Dragos to ensure that security tools do not inadvertently cause physical damage or service disruption.

Sources