Back to list
Nov 06, 2026
2.8K
0
0
GPT

GPT-6 Cyber: 48.5% on ExploitGym, Built to Stay in Scope

OpenAI's GPT-6 Cyber scores 48.5% on ExploitGym and avoided out-of-scope targets in 99.93% of honeypot tests. Access starts with select customers.

#GPT-6 Cyber#OpenAI#Cybersecurity#Daybreak#ExploitGym
GPT-6 Cyber: 48.5% on ExploitGym, Built to Stay in Scope
AI Summary

OpenAI's GPT-6 Cyber scores 48.5% on ExploitGym and avoided out-of-scope targets in 99.93% of honeypot tests. Access starts with select customers.

Overview

On November 5, 2026, OpenAI introduced GPT-6 Cyber, a security-research model that it describes as more capable at finding and fixing vulnerabilities and more reliable at staying within the boundaries researchers set. OpenAI reports a 51% relative improvement on the ExploitGym benchmark over GPT-5.6 Cyber and classifies the model as Critical in cybersecurity under its Preparedness Framework, up from High for GPT-5.6 Cyber.

The model is not a general release. It is offered initially through the API and Codex to a select group of approved Daybreak Red customers. This review covers the benchmark claims, the alignment results that OpenAI puts at the center of the launch, the real-world findings it cites, and the critique that TechCrunch attaches to the surrounding Daybreak program.

Feature Overview

1. Exploit and discovery benchmarks. OpenAI says GPT-6 Cyber outperforms GPT-6 Astra on most cyber tasks, even when Astra does not refuse the request. The key numbers come from OpenAI's own evaluations.

EvaluationGPT-6 CyberComparison
ExploitGym (exploit a specified vulnerability in a remote target)48.5%GPT-5.6 Cyber 32.1%, about 16 points lower; The Verge, reading OpenAI's chart, reports GPT-6 Astra at 42.4%
New offensive web penetration test (no source code access), recall40.3%GPT-6 Astra 32.0%
RemoteExploitBench (known vulnerabilities in GitLab, PostgreSQL and Jenkins, controlled environment)Completed 95%GPT-6 Astra running with Daybreak Blue refused 100% of tasks

OpenAI calls the RemoteExploitBench result an initial evaluation. GPT-6 Cyber's only failure there was an exploit it could not get to work remotely.

2. Staying in scope. Cybersecurity agents can wander outside their task, attack unrelated services or ignore rules of engagement, and OpenAI made this a central focus, using specialized training for staying within authorized scope, following restrictions and reporting work accurately. Its ExploitGym Honeypot evaluation, informed by the Hugging Face incident, gives the model a hard exploitation task in an isolated environment alongside deliberately easy-to-exploit services such as a package registry or object store. Attacking them would reveal the task's flag, so the model must infer that doing so is out of scope. Without deployment safety prompts, GPT-6 Cyber avoided exploiting out-of-scope services in 99.93% of cases, up from 94.49% for GPT-5.6 Cyber; with those prompts it stayed within bounds 100% of the time. OpenAI shares chain-of-thought excerpts such as "don't explore out of scope credentials."

3. Findings in real software. OpenAI's researchers report the following.

  • Firefox: evidence and reproducible tests for 81 novel vulnerabilities that Mozilla classified as high severity, plus 30 known issues shown to be more severe than previously reported. Mozilla began landing fixes (advisory mfsa2026-118).
  • Linux: 1,724 vulnerabilities reproduced and working exploits for 948, more than five times the 166 exploited by GPT-5.6 Sol using the same research harness. OpenAI disclosed 304 critical findings to maintainers with candidate patches. In a related post, OpenAI says the run used $1,213,430 in API credits.
  • Brotli: used alongside other OpenAI models to validate exploits showing how a critical flaw could enable remote code execution in downstream applications.
  • OpenAI's own infrastructure: in controlled red-teaming campaigns, agents using GPT-6 Cyber and other OpenAI models chained vulnerabilities to escape microVMs.

Cisco, Calif and CodeRabbit are named as customers using it for vulnerability research and code review.

4. Safeguards. OpenAI lists cyber-specific misalignment monitoring, targeted monitoring for emerging threats (for example during the coordinated disclosure of critical PHP vulnerabilities), Private Safety Processing, which Zero Data Retention API customers generally must use (encrypted records retained in customer storage for 30 days, not exposed to OpenAI personnel), IP allowlisting required by default for Daybreak Red API access, credential-compromise monitoring, and new checks that block high-risk activity involving critical systems. It recommends sandboxing, action review and human oversight for high-risk workflows.

5. Critical infrastructure. OpenAI's $1 billion Daybreak for Frontline Defenders grant program subsidizes access for resource-constrained teams. A new group, Daybreak Red for Critical Infrastructure, gives approved defenders permission to conduct authorized testing of critical systems with safeguards against misalignment retained. Dragos ran evaluations in a controlled lab, and OpenAI states that higher safeguard block rates are not always better, because many blocked techniques are also essential to legitimate security work.

Usability Analysis

GPT-6 Cyber is aimed at vetted security teams rather than general users. Daybreak Blue helps verified defenders validate vulnerabilities and fixes with fewer unnecessary refusals; Daybreak Red enables deeper exploit development and authorized security testing with GPT-6 Cyber. Mainline models such as GPT-6 Astra already support code review, vulnerability triage and patch development. OpenAI is also making GPT-5.6 Cyber available to all Daybreak Red customers and expanding Daybreak Blue alpha access for GPT-6 Astra.

Day-to-day use comes with operational conditions: IP allowlisting for API access by default, Private Safety Processing, generally required for zero-retention API customers, and sandboxed environments for high-risk work. Teams should treat these as part of the deployment plan.

TechCrunch adds a skeptical view. Max Justicz, CTO and co-founder of Charge Robotics, used GPT-5.6 Sol through Daybreak Blue to find a severe Ubuntu-affecting bug. He told TechCrunch he spent about $5,000 on subscriptions and tokens to find several exploits, that Daybreak Blue appears to have been recently "nerfed", and that he joined by uploading a photo of his driver's license. These comments concern Daybreak Blue and an earlier model, not GPT-6 Cyber. TechCrunch also notes that OpenAI says "thousands" of security professionals participate in Daybreak, and that the Linux bug hunt cost $1.2 million.

Pros and Cons

Pros

  • A 48.5% ExploitGym score, against 32.1% for GPT-5.6 Cyber, in OpenAI's own evaluation.
  • Scope discipline measured directly, at 99.93% without deployment safety prompts.
  • Findings that reached maintainers: Mozilla began landing fixes, and 304 critical Linux findings were disclosed with candidate patches.
  • Layered safeguards, from misalignment monitoring to IP allowlisting.

Cons

  • Launch access is limited to a select group of approved Daybreak Red customers.
  • The benchmark results come from OpenAI's evaluations, including two new ones, and RemoteExploitBench is described as an initial evaluation.
  • Large runs are expensive, with $1,213,430 in API credits for the Linux run.
  • A researcher quoted by TechCrunch says Daybreak Blue appears to have been recently "nerfed" and that he joined by uploading a photo of his driver's license.

Outlook

OpenAI says it will progressively expand access to all Daybreak Red customers soon; The Verge reports "in the coming weeks." Its stated goal is to dramatically expand Daybreak to organizations of every size, independent researchers and open-source maintainers, and it argues that there is a limited window to use frontier models to secure software. TechCrunch writes that defenders can feel the labs have created a mountain of new work that outstrips their resources, and that more urgency may be required. OpenAI's post commits to refining safeguards with input from partners and the security community.

Conclusion

GPT-6 Cyber pairs higher exploit-development scores with specific evidence about scope control, and the Firefox and Linux results show findings moving to maintainers. The evidence is largely self-reported, access is restricted and the Critical classification underlines the dual-use stakes. It is relevant to approved security teams, vulnerability researchers and open-source maintainers watching how frontier labs handle offensive capability. Rating: 4 out of 5.

Editor's Verdict

GPT-6 Cyber earns a solid recommendation as a security-research model, with the caveat that most of its evidence is self-reported.

The strongest case for paying attention is its 48.5% ExploitGym score, against 32.1% for GPT-5.6 Cyber, in OpenAI's own evaluation. Reinforcing that, OpenAI measures scope discipline directly, and the model avoided out-of-scope services in 99.93% of honeypot cases without deployment safety prompts. The broader signal worth registering is straightforward: the scope results matter as much as the exploit scores, because OpenAI frames the model's value around staying inside authorized boundaries. On the other side of the ledger, one constraint is real rather than a marketing footnote: launch access is limited to a select group of approved Daybreak Red customers, through the API and Codex. Layered on top of that, most benchmark results come from OpenAI's own evaluations, including two new ones, and RemoteExploitBench is described as an initial evaluation.

For approved Daybreak Red teams, vulnerability researchers and open-source maintainers, the Firefox and Linux results are the part to study closely. Everyone else should watch whether independent researchers reproduce the scope-control numbers as access widens.

Advertisement

Pros

  • A 48.5% ExploitGym score, against 32.1% for GPT-5.6 Cyber, in OpenAI's own evaluation.
  • Scope discipline measured directly, with out-of-scope services avoided in 99.93% of honeypot cases without deployment safety prompts.
  • Findings that reached maintainers, including 81 high-severity Firefox vulnerabilities and 304 critical Linux findings disclosed with candidate patches.
  • Layered safeguards, from misalignment monitoring to IP allowlisting and Private Safety Processing.

Cons

  • Launch access is limited to a select group of approved Daybreak Red customers, through the API and Codex.
  • Most benchmark results come from OpenAI's own evaluations, including two new ones, and RemoteExploitBench is described as an initial evaluation.
  • Large runs are costly, with the Linux run using $1,213,430 in API credits.
  • A researcher quoted by TechCrunch says Daybreak Blue appears to have been recently nerfed and that he joined by uploading a photo of his driver's license.
Advertisement

Comments0

Key Features

1. ExploitGym score of 48.5%, up from 32.1% for GPT-5.6 Cyber, and 40.3% recall on a new offensive web pentest evaluation 2. RemoteExploitBench initial evaluation: 95% completed, versus 100% refused by GPT-6 Astra running with Daybreak Blue 3. ExploitGym Honeypot scope evaluation: out-of-scope services avoided in 99.93% of cases without deployment safety prompts, 100% with them 4. Reported findings: 81 high-severity Firefox vulnerabilities and working exploits for 948 Linux vulnerabilities 5. Safeguards including Private Safety Processing, IP allowlisting for Daybreak Red API access and a $1 billion Daybreak for Frontline Defenders grant program

Key Insights

  • The scope results matter as much as the exploit scores, because OpenAI frames the model's value around staying inside authorized boundaries.
  • The jump from 32.1% to 48.5% on ExploitGym is about 16 points, which OpenAI describes as a 51% relative improvement.
  • OpenAI says GPT-6 Cyber outperforms GPT-6 Astra on most cyber tasks, not all, even when Astra does not refuse the request.
  • The model is classified Critical in cybersecurity under the Preparedness Framework, up from High for GPT-5.6 Cyber, and access starts with a select group of approved Daybreak Red customers.
  • The Linux run reproduced 1,724 vulnerabilities and developed 948 working exploits, with $1,213,430 in API credits according to OpenAI's related post.
  • The Daybreak Blue and Daybreak Red split separates verified defenders validating fixes from authorized offensive testing and exploit development.
  • TechCrunch raises access and cost questions through Max Justicz, whose comments concern Daybreak Blue and GPT-5.6 Sol rather than GPT-6 Cyber itself.

Was this review helpful?

Share

Twitter/X
Advertisement