Key Notes
- OpenAI has introduced GPT-6 Cyber for selected approved Daybreak Red customers through the API and Codex, classifying its cybersecurity capabilities as Critical.
- The model powered research that generated 948 Linux privilege escalation exploits, while OpenAI also reports 81 previously unknown high-severity Firefox vulnerabilities.
- Its reported benchmark gains and improved scope compliance come from specific evaluations, and the separate power-plant findings in the linked Dragos report concern earlier models.
OpenAI has released GPT-6 Cyber, a specialized model that can find software vulnerabilities and turn them into working exploits. The company rates its cybersecurity capabilities as Critical under its Preparedness Framework and is limiting the initial rollout to a select group of approved security customers.
The November 5 announcement describes a model built for authorized vulnerability research, penetration testing and patch development. It is available through the API and Codex to selected Daybreak Red customers, with wider access within that program planned. This is a restricted security release, rather than a general ChatGPT rollout.
Linux and Firefox Show What the Model Can Find
The Linux results give the release a concrete scale. In the research pipeline we covered earlier, agents reproduced 1,724 vulnerabilities and generated 948 working local privilege escalation exploits. Those attacks demonstrate how a user with limited access to a vulnerable system could gain root privileges; they do not represent 948 remote breaches of unrelated servers.
OpenAI’s Linux report puts the API cost at about $1.2 million. The release now identifies GPT-6 Cyber as the model behind that work. Its 948 exploits compare with 166 from GPT-5.6 Sol using the same research harness, a gain of more than 5x. Human researchers reviewed findings and candidate patches before reporting them to maintainers.
For Firefox, OpenAI reports 81 previously unknown vulnerabilities classified as high severity by Mozilla, plus evidence that 30 known issues were more serious than previously assessed. The company supplied reproducible tests and proposed fixes. These findings extend the work beyond proving that a model can solve a prepared security challenge.
Mozilla’s Firefox 159 advisory, published October 27, documents released fixes and credits OpenAI researchers and Codex Security for numerous high-severity bugs. It provides a record of remediation, although the announcement does not establish that every reported finding has already been patched. A discovered vulnerability, a demonstrated exploit and a released fix are separate stages of the work.
What the Exploit Benchmarks Measure
GPT-6 Cyber scored 48.5% on ExploitGym, compared with 32.1% for GPT-5.6 Cyber, according to OpenAI. The test asks a model to exploit a specified vulnerability in a remote target. That is a 16.4 percentage-point improvement, or about 51% in relative terms, within this evaluation.
On RemoteExploitBench, it completed 95% of tasks involving known vulnerabilities in applications such as GitLab, PostgreSQL and Jenkins in a controlled environment. GPT-6 Astra under Daybreak Blue refused all the tasks in the comparison. That refusal result is different from attempting an exploit and failing, so the comparison reflects access rules as well as technical ability. The 95% figure is not a success rate against arbitrary internet-facing servers.
A Stronger Agent Still Needs Boundaries
The system card identifies GPT-6 Cyber as a security-specialized version of GPT-6 Astra. Its training aims to reduce refusals for authorized security work while preserving the limits of a user’s permission. Critical is OpenAI’s capability classification, up from High for GPT-5.6 Cyber, rather than a claim that the model has been released without controls.
In a honeypot evaluation, the agent could complete its task by attacking a supporting service it was not authorized to exploit. It stayed within scope in 99.93% of tests without deployment safety instructions and 100% with them. Those are observed evaluation results, not a guarantee that every future agent session will remain within bounds.
Other tests show limits. External evaluator Irregular found no successful attacks on fully hardened targets in its evaluation, and neither GPT-6 Cyber nor Astra solved its 15 hardest challenges. OpenAI is pairing access restrictions with monitoring, account protections and additional checks for high-risk activity. Its guidance retains isolated environments, explicit authorization and human oversight as part of deployment.
The Power-Plant Finding Comes From Earlier Research
OpenAI also highlights industrial-security work with Dragos, including two previously unknown flaws found in engineering software on a mock power-plant network. The linked Dragos report is dated September 2026 and names GPT-5.5 Cyber and GPT-5.6 Sol as the evaluated models. It should therefore be read as earlier research informing safeguards, rather than a standalone GPT-6 Cyber result.
The flaws exposed sensitive engineering records, including protective-relay settings and breaker logic. Dragos says the power-range assessment did not test a disruptive or destructive attack. The work demonstrates how AI can help map industrial systems and uncover weaknesses, while leaving a clear distinction between discovering access to engineering data and taking control of a real power station.
Disclaimer: AIstify is an independent media brand owned and operated by NuvexMedia LLC, publishing news, research, and insights on artificial intelligence, emerging technologies, automation, and related industries. NuvexMedia LLC invests in and collaborates with companies across the AI, technology, software, and digital innovation sectors. These relationships do not influence AIstify’s editorial coverage, and the publication maintains full editorial independence to provide accurate, timely, and objective information. © 2026 NuvexMedia LLC. All rights reserved. This content is for informational purposes only and should not be considered legal, tax, investment, financial, or other professional advice.
