Meet GPT-6 Cyber: A Security Model That Finds and Fixes Vulnerabilities, and Stays in Scope

Hi, it's Shii-chan!

OpenAI News openai.com

A new model for security work? I leaned right in. Let me tell you about it.

What was announced?

In OpenAI News, GPT-6 Cyber was announced. It is a cybersecurity model for finding exploitable vulnerabilities and developing fixes, trained to stay within authorized scope.

Why it matters

A security research agent can drift outside its assigned task, attack unrelated services, or ignore the researcher's rules of engagement. GPT-6 Cyber was trained with alignment as a central focus, alongside better vulnerability research.

What changes

• You can hand over vulnerability discovery, exploit development, and patch drafting with higher accuracy
• Better scope-keeping makes autonomous work easier to trust
• OpenAI says it wants to dramatically expand its Daybreak program for defenders

Dive Deep

Benchmarks

• ExploitGym (exploiting a specified vulnerability on a remote target): 48.5%, about 16 points above GPT-5.6 Cyber's 32.1%
• A new offensive web penetration testing evaluation (no source code access): 40.3% recall of vulnerabilities, versus 32.0% for GPT-6 Astra
• RemoteExploitBench (known vulnerabilities in apps such as GitLab, PostgreSQL, and Jenkins in a controlled environment): GPT-6 Astra with Daybreak Blue refused every task, while GPT-6 Cyber completed 95%. Its only failure was one exploit it couldn't get working remotely
• Under the Preparedness Framework, its cybersecurity classification moved from High to Critical

Staying in scope

• ExploitGym Honeypot places easy-to-exploit supporting services, such as a package registry, in the environment and checks whether the model infers that attacking them is out of scope
• Without deployment prompts, it avoided out-of-scope services in 99.93% of cases (GPT-5.6 Cyber: 94.49%). With them, 100%

Real-world results (by OpenAI researchers)

• Firefox: evidence and reproducible tests for 81 novel vulnerabilities that Mozilla classified as high severity, and 30 known issues shown to be more severe than reported
• Linux kernel: 1,724 vulnerabilities reproduced and working exploits for 948, versus 166 for GPT-5.6 Sol in the same harness. 304 critical findings were disclosed to maintainers with candidate patches
• Security teams at Cisco, Calif, and CodeRabbit use it for vulnerability research and code review

Misuse safeguards

• Cyber-specific misalignment monitoring that checks whether actions match the user's authorization
• Private Safety Processing: API customers using Zero Data Retention generally must use it. It reviews encrypted records kept in customer storage for 30 days without exposing prompts and responses to OpenAI personnel
• IP allowlisting is required by default for Daybreak Red API access
• New checks block high-risk activity involving critical systems. Sandboxing, action review, and human oversight are recommended

The full results and evaluation settings are in the system card the announcement links to.

Wrap-up

• GPT-6 Cyber is a security model for finding vulnerabilities, developing exploits, and fixing them
• 48.5% on ExploitGym and 95% completion on RemoteExploitBench are big jumps from the previous generation
• It avoided out-of-scope services in 99.93% of tests without deployment prompts
• Results include 81 Firefox vulnerabilities and 948 Linux exploits
• Safeguards include monitoring, IP allowlisting, and Private Safety Processing

This is for security researchers, open-source maintainers, and defensive teams!