OpenAI releases GPT-6 Cyber, a security model it says stays in scope
OpenAI says GPT-6 Cyber scores 48.5% on ExploitGym, up from 32.1% for GPT-5.6 Cyber, and stayed in scope in 99.93% of honeypot tests without safety prompts. These are OpenAI's own results.
If you do security research or code review, this shows what OpenAI's Daybreak tiers offer verified defenders. The figures are OpenAI's own, so test the model on your own codebase before relying on them.
What happened
OpenAI introduced GPT-6 Cyber on November 5, 2026, in its own announcement. The model is built to find and validate exploitable vulnerabilities and to help write fixes. OpenAI says it was trained to stay inside the scope a researcher sets.
Every figure below is OpenAI's claim, taken from its post and system card. The post shows no independent verification.
Key details
- Linux: The model reproduced 1,724 vulnerabilities and built working exploits for 948. GPT-5.6 Sol exploited 166 with the same research harness. OpenAI disclosed 304 critical findings to maintainers with candidate patches.
- Firefox: OpenAI researchers produced evidence and reproducible tests for 81 novel vulnerabilities that Mozilla classed as high severity. Mozilla has released fixes for many of them, according to OpenAI.
- Risk rating: OpenAI classifies the model as Critical in cybersecurity under its Preparedness Framework. GPT-5.6 Cyber was rated High.
- Users named by OpenAI: Cisco, Calif and CodeRabbit use it for vulnerability research and code review.
Benchmarks at a glance
| Test | GPT-6 Cyber | Comparison (OpenAI's numbers) |
|---|---|---|
| ExploitGym (exploit a specified vulnerability) | 48.5% | GPT-5.6 Cyber: 32.1% |
| Web pentesting (recall, no source code) | 40.3% | GPT-6 Astra: 32.0% |
| RemoteExploitBench (tasks completed) | 95% | GPT-6 Astra with Daybreak Blue refused 100% of tasks |
| ExploitGym Honeypot, no deployment prompt (stayed in scope) | 99.93% | GPT-5.6 Cyber: 94.49% |
| ExploitGym Honeypot, with deployment prompt | 100% | Not stated |
OpenAI says its only RemoteExploitBench miss was an exploit it could not get to work remotely. The honeypot test places easy-to-attack services, such as a package registry, next to a hard task. The model has to infer that attacking them is out of scope.
Access through Daybreak
OpenAI ties the model to its Daybreak programme. Daybreak Blue helps verified defenders validate vulnerabilities and fixes with fewer unnecessary refusals. Daybreak Red covers deeper exploit development and authorized security testing with GPT-6 Cyber. Mainline models such as GPT-6 Astra already handle code review, triage and patch development.
Why it matters
OpenAI names going out of scope as a key risk for security agents: hitting unrelated services or ignoring the researcher's rules. It published a specific test for that and says it paired the model with stronger misuse detection.
The benchmark results are self-reported. Mozilla landing fixes is a visible outside signal for the Firefox work.
FAQ
What is GPT-6 Cyber? It is an OpenAI model for finding and validating exploitable vulnerabilities and developing fixes. OpenAI says it was trained to stay within authorized scope.
How do I get access to GPT-6 Cyber? The post points to Daybreak. Daybreak Red enables exploit development and authorized security testing with GPT-6 Cyber. The source excerpt gives no signup steps.
Is GPT-6 Cyber better than GPT-5.6 Cyber? OpenAI says so on its own benchmarks, including ExploitGym at 48.5% versus 32.1%. The announcement includes no independent results.
Data: OpenAI · chart by OpenBeat
Checked claims
| Claim | Status | Evidence |
|---|---|---|
| OpenAI says GPT-6 Cyber scores 48.5% on ExploitGym, up from 32.1% for GPT-5.6 Cyber. | confirmed | GPT‑6 Cyber scores 48.5%, up from 32.1% for GPT‑5.6 Cyber |
| In the honeypot scope test without deployment prompts, OpenAI says the model avoided out-of-scope services in 99.93% of cases, versus 94.49% for GPT-5.6 Cyber. | confirmed | it avoided exploiting out-of-scope services in 99.93% of cases, up from 94.49% for GPT‑5.6 Cyber |
| OpenAI says it produced evidence and reproducible tests for 81 novel Firefox vulnerabilities that Mozilla classed as high severity. | confirmed | GPT‑6 Cyber produced evidence and reproducible tests for 81 novel vulnerabilities classified by Mozilla as high severity |
| OpenAI says the model reproduced 1,724 Linux vulnerabilities and developed working exploits for 948. | confirmed | GPT‑6 Cyber reproduced 1,724 vulnerabilities and developed working exploits for 948 |
| OpenAI classifies GPT-6 Cyber as Critical in cybersecurity under its Preparedness Framework, up from High for GPT-5.6 Cyber. | confirmed | we classify the model as Critical in cybersecurity under our Preparedness Framework, up from High for GPT‑5.6 Cyber |
| On RemoteExploitBench, GPT-6 Astra with Daybreak Blue refused every task while GPT-6 Cyber completed 95%. | confirmed | GPT‑6 Astra running with Daybreak Blue refused 100% of tasks , while GPT‑6 Cyber completed 95% |
Question for readersIf you run security reviews on your own code, which task would you hand to an agent like this first, and which would you keep manual?
Discussion · 0
No comments yet. Start the discussion.
Sign in to join the discussion.