OpenAI says its new security model carries its highest cyber risk rating and that an automated pipeline built exploits for 948 Linux flaws. Almost every number comes from OpenAI itself.
CTRL + ALT News | Artificial Intelligence | Information checked on 8 November 2026
A note on transparency: Ctrl + Alt News uses Anthropic’s Claude in its work, and OpenAI is a competitor of Anthropic. We have kept to what is reported, attributed the numbers to the companies that gave them and marked what we could not check.
Two OpenAI announcements landed within days of each other. On 5 November 2026 the company released GPT-6 Cyber, a model for offensive security work that is open only to approved customers. And a report dated 4 November, “A thousand fewer paths to root”, describes an experiment in which an automated pipeline built working exploits for 948 flaws in the Linux kernel.
We could not open OpenAI’s own pages, which returned an access error. What follows comes from two outlets, CSO Online and Implicator.ai, and the most important caveat is simple: the performance numbers are OpenAI’s own, not independently verified.
What GPT-6 Cyber is, and who can use it
According to Implicator.ai, GPT-6 Cyber is available only to approved customers of OpenAI’s “Daybreak Red” programme, the tier for authorised offensive work such as developing exploits and penetration testing, through the API and Codex. API requests are restricted to approved networks by default. OpenAI said access should widen to all Daybreak Red customers “in the coming weeks”, the outlet reports.
OpenAI classes the model as “Critical” for cyber security under its Preparedness Framework, the highest tier and one above the “High” given to the earlier GPT-5.6 Cyber, Implicator.ai says. In plain terms: OpenAI itself regards this as its riskiest cyber model so far, which is why access is limited.
The test numbers, and the caveats
Implicator.ai reports figures from OpenAI’s own evaluations:
- In a “honeypot” test, where a model works in an environment with off-limits services, GPT-6 Cyber used services outside its scope in 0.07% of runs when no deployment safety instructions were given, against 5.51% for GPT-5.6 Cyber. With the instructions, the rate was zero. OpenAI did not describe what happened in the 0.07% of runs.
- On ExploitGym, a test of exploiting a specified vulnerability, GPT-6 Cyber scored 48.5% against 32.1% for GPT-5.6 Cyber.
Two cautions from the same report. First, the drop from 5.51% to 0.07% compares two different models; it is not one model improving. Second, these are OpenAI’s evaluations, and the company has not described behaviour in real deployments. A spokesperson, Fouad Matin, is quoted saying that “evaluation awareness”, a model recognising it is being tested, is a concern. Implicator.ai also says the page that carries these figures is OpenAI’s own system card, which we have not read.
The Linux report: 948 exploits
OpenAI’s report, dated 4 November, describes an experiment run in October. According to CSO Online, an automated pipeline found vulnerabilities, validated them, built exploits and generated patches. It “created exploits for 948 distinct local privilege-escalation vulnerabilities” in the Linux kernel.
“Local privilege escalation” means an attacker who already has a foothold on a machine gains higher rights, up to administrator (“root”) level. It is a step in an attack, not a way in from outside.
What CSO Online reports:
- Cost: $1,213,430 in API credits, about $1,279 per exploit.
- Human review: researchers review the root-cause analysis, proofs of concept and patches before anything is sent to the maintainers.
- Status of fixes: during the campaign, 61 of the flaws already had upstream patches, 43 had patches in staging and 299 were in public discussion. The report does not give a full picture of how many of the 948 have been fixed.
- Scale: the pipeline found five times as many unique privilege-escalation flaws as an earlier experiment with GPT-5.6 Sol, and the agents used at least 10 new exploitation techniques in a review of 166 exploits.
Caveats OpenAI itself gives
CSO Online says OpenAI reports that in an early experiment, 20 of 200 “successful” exploits gained root through a different zero-day flaw from the intended one, so checks were added. The early AI-written patches sometimes included unnecessary changes or missed variants.
OpenAI also estimates the exploits’ market value at $18.9 million to $66.3 million. CSO Online stresses that this compares them with prizes at hacking contests such as Pwn2Own and TyphoonPWN and “does not represent actual bounty payouts”.
The two outlets also differ on some counts. Implicator.ai says the system reproduced 1,724 vulnerabilities and reported 304 critical findings, figures CSO Online does not give. CSO Online cites hundreds of reports and patches and the 61, 43 and 299 above. We cannot reconcile these without OpenAI’s report.
Why it matters, and why to be careful
CSO Online quotes an analyst from IDC, Sakshi Grover, on the gap between finding a flaw and having it fixed in live systems. That is the real worry: if AI finds flaws faster than fixes reach machines, defenders have less time. It echoes the “race to shorten patch deployment timelines” the NCSC describes in its 2026 review.
For readers, the takeaway is modest. Linux runs on a vast range of systems, and nothing in these reports names a flaw you must act on today. Keeping software updated is the standard advice. We showed what happens when serious flaws are attacked before patching in the PHP2Shell flaws, and how AI tooling can be exposed in the OpenCode breach.
Frequently asked questions
Can anyone use GPT-6 Cyber? No. According to Implicator.ai, only approved Daybreak Red customers can, through the API and Codex.
Did AI really find 948 Linux flaws? CSO Online reports, from OpenAI’s report, that the pipeline created exploits for 948 distinct local privilege-escalation flaws. Nobody outside OpenAI has verified that, and OpenAI’s report was not accessible to us.
Is the cost real? The $1,213,430 is OpenAI’s reported API cost for the campaign. The $18.9 to $66.3 million “value” is a hypothetical comparison, not money paid.
Is the 0.07% an improvement? It is lower than the 5.51% of another model. That compares two models, and OpenAI has not explained what happened in the 0.07% of runs.
What we could not confirm
- OpenAI’s own pages: the GPT-6 Cyber announcement, its system card and the Linux report.
- Any independent check of the honeypot and ExploitGym numbers.
- What happened in the 0.07% of honeypot runs.
- The reconciliation of the counts that differ between CSO Online and Implicator.ai.
- Whether GPT-6 Cyber is available in the UK and on what terms.
Sources: CSO Online, 5 November 2026; Implicator.ai, 5 November 2026.
Leave a Reply