FUNDING · MODELS · #3157
OpenAI launches GPT-6 Cyber, a 'Critical' cybersecurity model with stronger safeguards
OpenAI introduced GPT-6 Cyber, a model it classifies as 'Critical' under its Preparedness Framework; in internal evaluations (ExploitGym HoneyPot) it scored 48.5% on exploit benchmarks—outperforming GPT-5.6 Cyber (32.1%) and slightly beating GPT-6 Astra (42.4%)—and reportedly helped find 81 previously unknown Firefox vulnerabilities. OpenAI says GPT-6 Cyber goes out of bounds under 1% of the time (vs ~5% for GPT-5.6 Cyber), includes new checks to block high-risk activity, and is rolling out to select members of its Daybreak Red program for authorized defensive cybersecurity work.
KEY POINTS
- OpenAI introduced GPT-6 Cyber, a model it classifies as 'Critical' under its Preparedness Framework; in internal evaluations (ExploitGym HoneyPot) it scored 48.5% on exploit benchmarks—outperforming GPT-5.6 Cyber (32.1%) and slightly beating GPT-6 Astra (42.4%)—and reportedly helped find 81 previously unknown Firefox vulnerabilities.
- OpenAI says GPT-6 Cyber goes out of bounds under 1% of the time (vs ~5% for GPT-5.6 Cyber), includes new checks to block high-risk activity, and is rolling out to select members of its Daybreak Red program for authorized defensive cybersecurity work.
- This matters because a 'Critical'-rated model that finds real-world zero‑day exploits while claiming stronger guardrails changes the balance of defensive capability and misuse risk in AI cybersecurity.
DECISION BRIEF
CONFIDENCE
WHAT CHANGED
OpenAI announced GPT‑6 Cyber, a cybersecurity-focused model that OpenAI classifies as “Critical” and that it says is trained to stay within authorized scope. OpenAI reports internal evaluation results (ExploitGym HoneyPot) with GPT‑6 Cyber scoring 48.5% on exploit benchmarks (vs. 32.1% for GPT‑5.6 Cyber and 42.4% for GPT‑6 Astra), reporting it goes “out of bounds” under 1% of the time (vs ~5% for GPT‑5.6 Cyber). OpenAI and reporting outlets state the model helped find 81 previously unknown Firefox vulnerabilities and that OpenAI is making the model available to select members of its Daybreak Red program for authorized defensive work.
WHY NOW
Sources report GPT‑6 Cyber is both more capable on exploit benchmarks and is classified as “Critical” (a class OpenAI links to the ability to identify and develop functional zero‑day exploits). Media coverage and OpenAI’s own statements highlight recent rapid increases in model capability, examples of AI‑assisted vulnerability findings, and a perceived narrow window for defenders to act. Those facts together imply a near‑term change in the balance between defensive capability and potential misuse risk.
WHO IS AFFECTED
Directly named in sources: OpenAI’s Daybreak programs (Daybreak Red and earlier Daybreak Blue), independent/third‑party defensive researchers granted Daybreak access, open‑source maintainers and projects cited (Firefox and the Linux/Debian code referenced in reporting), and organizations responsible for hardened critical systems (per OpenAI’s Critical classification).
CONFIRMED
- OpenAI announced GPT‑6 Cyber as a cybersecurity model and says it is trained to stay within authorized scope. (OpenAI excerpt) - OpenAI classifies GPT‑6 Cyber as “Critical” under its Preparedness Framework. (The Verge) - In OpenAI’s internal ExploitGym HoneyPot evaluation, GPT‑6 Cyber scored 48.5% on exploit benchmarks; GPT‑5.6 Cyber scored 32.1% and GPT‑6 Astra scored 42.4%. (The Verge) - OpenAI reports GPT‑6 Cyber went out of bounds under 1% of the time in its testing, versus ~5% for GPT‑5.6 Cyber. (The Verge) - OpenAI and reporting outlets say the model helped find 81 previously unknown vulnerabilities in Firefox. (The Verge; TechCrunch) - OpenAI is rolling out the model to select members of its Daybreak Red program for authorized defensive cybersecurity work; earlier Daybreak programs (Daybreak Blue) have been used to give defenders access to frontier models. (OpenAI excerpt; TechCrunch)
UNCERTAIN
- All capability and safety metrics reported are from OpenAI’s internal evaluations or company statements; independent third‑party replication of ExploitGym HoneyPot scores and the “<1% out‑of‑bounds” rate is not provided in the supplied excerpts. (The Verge notes concerns about evaluation awareness.) - The degree to which the reported findings (e.g., 81 Firefox issues; 948 potential Linux kernel issues mentioned in reporting) correspond to validated, publicly disclosed CVEs or to exploitable zero‑day exploits is not shown in the supplied excerpts. (TechCrunch claims 81 novel Firefox vulnerabilities and 948 potential Linux kernel issues; the excerpts do not supply external verification.) - It is unclear how representative internal test scenarios (ExploitGym HoneyPot) are of live‑production adversarial conditions because evaluation awareness and potential test‑aware behavior are explicitly noted. (The Verge) - The excerpts include slightly different product naming/phrasing across outlets (e.g., TechCrunch references “ChatGPT Cyber” and earlier “GPT 5.6 Sol” usage), creating minor inconsistencies in public reporting versus OpenAI messaging. The supplied excerpts do not reconcile those naming differences.
WHAT TO WATCH
Observable signals to watch in supplied public sources: - Independent third‑party evaluations or audits reproducing ExploitGym HoneyPot results or publishing their own exploit‑finding benchmarks. (look for published reports or preprints) - Public disclosures, CVE entries, or vendor patches that trace attribution to findings from OpenAI models (e.g., Debian/Linux, Firefox advisories that cite OpenAI or Daybreak assistance). - OpenAI publishing detailed methodology, benchmark data, or safety‑evaluation artifacts for external review. (company blog or technical appendices) - Changes to Daybreak program access rules, participant lists, or scope (expansions beyond select Daybreak Red members). (OpenAI announcements or reporting) - Independent news or security‑research reports describing misuse, leakage, or successful offensive use of comparable models. (security blogs, vendor advisories, media follow‑ups)
WHY IT MATTERS
This matters because a 'Critical'-rated model that finds real-world zero‑day exploits while claiming stronger guardrails changes the balance of defensive capability and misuse risk in AI cybersecurity.
EVIDENCE MAP
4Editorial claims linked to specific sources, with support, contradiction and context shown separately.
OpenAI introduced GPT‑6 Cyber and classifies it as 'Critical' under its Preparedness Framework.
SUPPORTEDChecked 2026-11-05 · 1 supporting
The Verge excerpt states GPT‑6 Cyber is classified as 'Critical' under OpenAI's Preparedness Framework.
CONTEXT OpenAIMEDIA · 2026-11-05OpenAI's announcement introduces GPT‑6 Cyber as a cybersecurity model trained to stay within authorized scope (announcement excerpt).
In ExploitGym evaluations, GPT‑6 Cyber scored 48.5%; GPT‑5.6 Cyber scored 32.1%; GPT‑6 Astra scored 42.4%, with GPT‑6 Cyber outperforming the prior Cyber model and slightly outperforming Astra on these benchmarks.
SUPPORTEDChecked 2026-11-05 · 1 supporting
OpenAI says GPT‑6 Cyber goes 'out of bounds' less than 1% of the time (versus ~5% for GPT‑5.6 Cyber) and is trained/updated with checks to stay within authorized scope and block high‑risk activity.
SUPPORTEDChecked 2026-11-05 · 2 supporting
OpenAI's announcement excerpt describes GPT‑6 Cyber as trained to stay within authorized scope, indicating added safeguards in the model announcement.
SUPPORTS The Verge AIMEDIA · 2026-11-05The Verge excerpt reports the 'out of bounds' rates (<1% for GPT‑6 Cyber vs ~5% for GPT‑5.6 Cyber) and quotes OpenAI staff about improvements on this evaluation.
OpenAI reported GPT‑6 Cyber helped find 81 previously unknown vulnerabilities in Firefox and identified 948 potential issues while scanning/preparing patches for the Linux kernel.
SUPPORTEDChecked 2026-11-05 · 2 supporting
The Verge excerpt also states the model helped find 81 previously unknown vulnerabilities in Firefox and mentions scanning/preparing patches for the Linux kernel.
SUPPORTS TechCrunch AIMEDIA · 2026-11-05TechCrunch's excerpt reports that OpenAI said the model found 948 potential issues in the Linux kernel and 81 'novel' vulnerabilities in Firefox.
SOURCES & TIMELINE
4Introducing GPT‑6 Cyber, OpenAI’s cybersecurity model for finding exploitable vulnerabilities and developing fixes, trained to stay within authorized scope.
Max Justicz, the chief technology officer and cofounder of the startup Charge Robotics, used an OpenAI model to find a severe vulnerability that — if exploited — could allow hackers to take control of servers running Ubuntu, a widely-used open-source Linux based operating system. Justicz, who has a hobbyist’s interest in cybersecurity and previously found a vulnerability in the same codebase, told TechCrunch that he…
Classified as ‘Critical,’ GPT-6 Cyber outperforms GPT-6 Astra on cyber-related benchmarks. OpenAI is launching GPT-6 Cyber, the latest version of its cybersecurity AI model, which it says is better at finding vulnerabilities and less likely to go rogue and hack other platforms in the process. OpenAI says that during testing, GPT-6 Cyber went out of bounds less than 1 percent of the time, down from around 5 percent w…
OpenAI has unveiled GPT-6 Cyber, an AI model specifically trained to find security vulnerabilities, develop working exploits, and suggest patches. In targeted tests, GPT-6 Cyber found 81 previously unknown critical vulnerabilities in Firefox, which Mozilla has already patched. In the Linux kernel, the model reproduced 1,724 vulnerabilities and developed working exploits for 948 of them, more than five times as many…