Gemini 4 Carbon and GPT-6 Cyber, defenders first

Gemini 4 Carbon and GPT-6 Cyber, eight days apart

Google DeepMind announced Gemini 4 Carbon on 28 October 2026.¹¹1Koray Kavukcuoglu, Google, "Introducing Gemini 4 Carbon", 28 October 2026: Carbon described as Google's strongest model with "state-of-the-art cybersecurity and long horizon capabilities" and "a significant leap in cybersecurity capabilities"; Argon launched two weeks earlier; teams saw Carbon run for days; SEC-bench Pro 88.2%, ExploitGym 54.8% ("a new state-of-the-art"), CWE-bench 71.0%, all without guardrails; more than 2x secure code generation over Argon; Chrome team "more than 10X" claim against "another lab's large frontier model"; Wiz result without a number; first on the Artificial Analysis Intelligence Index v5.0 at 59; WebDev Arena +89 Elo, fourth at 1768; same safety stack as Argon; Terminal-Bench 4.0 cheating 0.0% and attempts 0.4%; high-severity misalignment alerts down almost 80%; Gray Swan IPI 0.0% against 0.7% for Argon; unsupported claims cut roughly 2x against Argon, overreach down 26% and under-service down 29%; "the U.S. government's voluntary process for pre-release model testing", agency not named; access for a subset of cyber defenders, then trusted partners through Fairwind, others "as soon as possible"; $2 and $10 per million tokens, cached input 95% off; later $4 and $20, with no end date. No Carbon model card found at the likely Google DeepMind addresses on 6 November 2026. https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/ OpenAI announced GPT-6 Cyber on 5 November 2026 and published its system card on the same day.²²2OpenAI, "Introducing GPT-6 Cyber", 5 November 2026: ExploitGym 48.5% against 32.1% for GPT-5.6 Cyber and "a 51% relative improvement"; web penetration test recall without source code 40.3% against 32.0% for GPT-6 Astra; RemoteExploitBench, Astra with Daybreak Blue refused 100% and GPT-6 Cyber completed 95%; outperforms Astra on most cyber tasks; Critical, up from High for GPT-5.6 Cyber; scope training and 99.93% avoidance without the deployment prompt, 100% with it; 948 Linux privilege escalation exploits; 81 high-severity Firefox vulnerabilities, many fixed in MFSA 2026-118; candidate patches; Cisco, Calif and CodeRabbit; use on OpenAI infrastructure; Astra, Daybreak Blue and Daybreak Red tiers; "dramatically expand Daybreak"; "a limited window to use frontier models to secure software"; no price. https://openai.com/index/introducing-gpt-6-cyber/ ³³3OpenAI, GPT-6 Cyber system card, Deployment Safety Hub, 5 November 2026: availability only to approved Daybreak Red users and organizations; ExploitGym 48.45% against 42.43% for Astra; ExploitGym Hardened 36.94% against 29.69%; SEC-Bench Pro pass@1 87.02% against 85.38%; SRE-Bench 92.84% against 87.98%; ExploitBench Internal Port 42% against 39%; Irregular results on CyScenarioBench (11 of 11 against 8 of 11, average 61%, up 7 points) and FrontierCyber (90 of 288 against 96 of 288, none of the 15 hardest, no success on fully hardened targets) and atomic challenges (22 of 22); "Like GPT-6 Astra, we classify GPT-6 Cyber as Critical in Cybersecurity under our Preparedness Framework"; High in biological and chemical, below High in AI self-improvement; no rationale for Critical, no Critical-specific safeguards, no mention of halting; ExploitGym honeypot 0.07% and 0%, Astra 0%, GPT-5.6 Sol 49.5%, GPT-5.6 Cyber 5.51%; Auto-Review 0% against 0.3% for Sol; warnings, stop 73.00% against 78.38%, continue 80.43% against 72.19%; harmful-outcome replays 28.40% against 26.70%; broken search tool errors 0.33%, 6.92% and 79.00%; Daybreak Red requirements for organisations and individuals; eligibility reviews; IP allowlisting by default; Private Safety Processing for Zero Data Retention customers; Codex app and API; new checks for industrial controls, essential services and election systems; misalignment monitoring and revocation; NIST CAISSI and Irregular assessments without results; no count of users or organisations. https://deploymentsafety.openai.com/gpt-6-cyber The two launches came 8 days apart.⁴⁴4Derived from the Carbon launch post (Google, 28 October 2026) and the GPT-6 Cyber launch post (OpenAI, 5 November 2026): 28 October to 5 November 2026 is 3 days in October plus 5 in November, or 8 days. https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/ https://openai.com/index/introducing-gpt-6-cyber/ Each lab gave its model first to vetted cyber defenders, and neither model is open to the general public.¹¹³³

Both labs report a score on ExploitGym, a benchmark of exploitation tasks. Google puts Carbon at 54.8%.¹¹ OpenAI puts GPT-6 Cyber at 48.5%.²² OpenAI also classifies GPT-6 Cyber as Critical in cybersecurity under its Preparedness Framework.³³ That framework defines a High level and a Critical level for cyber capability.⁵⁵5OpenAI, "Preparedness Framework", Version 2, last updated 15 April 2025: High and Critical cyber definitions; High responses on security controls, safeguards against misuse and misalignment, and contributions to cyberdefence policies and tools; Critical response "Until we have specified safeguards and security controls standards that would meet a Critical standard, halt further development". https://cdn.openai.com/pdf/18a02b5d-6b67-4cec-ab64-68cdfbddebcd/preparedness-framework-v2.pdf

Carbon followed Gemini 4 Argon closely. Argon became generally available on 13 October 2026, and Carbon was announced 15 days later.⁶⁶6Koray Kavukcuoglu, Google, "Introducing Gemini 4 Argon", general availability post, 13 October 2026: Argon generally available for developers and enterprises via the API; "After the introductory period expires on November 30, 2026," rates of $4 and $20 per million tokens. https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/ ⁷⁷7Derived from the Argon general availability post (Google, 13 October 2026) and the Carbon launch post (Google, 28 October 2026): 28 minus 13 is 15 days. https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/ https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/ The Carbon post itself describes Argon as launched two weeks earlier.¹¹ PARALLAX covered that staged release in an earlier explainer on Gemini 4 Argon.

What Gemini 4 Carbon is and how Google scores it

Gemini 4 Carbon is a Google DeepMind model that follows Gemini 4 Argon in the Gemini 4 line. Koray Kavukcuoglu announced it in a post on 28 October 2026, signing as senior vice president of Google DeepMind and chief AI architect of Google.¹¹ He describes Carbon as Google's strongest model, with "state-of-the-art cybersecurity and long horizon capabilities".¹¹

Kavukcuoglu writes that Carbon delivers "a significant leap in cybersecurity capabilities".¹¹ Google teams saw the model run on tasks for days.¹¹ Google has disclosed neither the architecture nor the parameter count.¹¹

Google reports one score for each of three cyber tasks. For finding vulnerabilities it uses SEC-bench Pro, where Carbon scores 88.2%.¹¹ On ExploitGym, its exploitation test, Carbon scores 54.8%, which Google calls "a new state-of-the-art".¹¹ For patching the post gives CWE-bench, at 71.0%.¹¹ Google ran all of these benchmarks without guardrails.¹¹

Some of Google's claims come without a public score. On internal benchmarks, Carbon more than doubles Argon's success rate at generating secure code.¹¹ The Chrome team reported that Carbon was more than 10X better at high quality patches than "another lab's large frontier model".¹¹ The post does not name that lab or its model. Wiz, a security company, observes Carbon leading on web penetration testing benchmarks. Google prints no number for that result.¹¹

Google also cites general rankings. Its post places Carbon first on the Artificial Analysis Intelligence Index v5.0, with a score of 59.¹¹ The Artificial Analysis page for the model, read on 6 November 2026, shows a different figure. It lists Gemini 4 Carbon at its High setting with 60 on index v5.0.1, ranked first of 69 models.⁸⁸8Artificial Analysis, Gemini 4 Carbon model page, undated, read on 6 November 2026: "Gemini 4 Carbon (High)" at 60 on the Artificial Analysis Intelligence Index v5.0.1, ranked first of 69; "median: 37". https://artificialanalysis.ai/models/gemini-4-carbon The same page gives a median of 37 among comparable models.⁸⁸ The two figures refer to different versions of the index.

The Artificial Analysis leaderboard on the same day placed Claude Opus 5.5 at its max setting second, with 59.⁹⁹9Artificial Analysis, models leaderboard, undated, read on 6 November 2026: Gemini 4 Carbon (high) 60, Claude Opus 5.5 (max) 59, Gemini 4 Carbon (medium) 58, Claude Opus 5.5 (xhigh) 57, GPT-6 Astra (max) 57; GPT-6 Cyber not listed. https://artificialanalysis.ai/leaderboards/models GPT-6 Astra from OpenAI scored 57 at its max setting.⁹⁹ On WebDev Arena, Carbon gained 89 Elo points and reached fourth place at 1768.¹¹ That puts the previous Gemini score at 1679.¹⁰¹⁰10Derived from the Carbon launch post (Google, 28 October 2026): 1768 minus 89 is 1679. https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/

What GPT-6 Cyber is and how its scores compare with Carbon's

GPT-6 Cyber is a model OpenAI built for cybersecurity work, announced on 5 November 2026.²² Its system card says it "is available only to approved Daybreak Red users and organizations."³³ OpenAI compares it with two earlier models. One is GPT-5.6 Cyber, the previous specialised model. The other is GPT-6 Astra.²²³³

Against GPT-5.6 Cyber, the launch post claims "a 51% relative improvement on ExploitGym".²² The underlying scores are 48.5% for GPT-6 Cyber and 32.1% for GPT-5.6 Cyber.²² The ratio of those two scores is 1.51, and the gap is 16.4 percentage points.¹¹¹¹11Derived from the GPT-6 Cyber launch post (OpenAI, 5 November 2026): 48.5 divided by 32.1 is 1.511, a relative improvement of 51%; 48.5 minus 32.1 is 16.4 points. https://openai.com/index/introducing-gpt-6-cyber/

The system card gives finer figures against GPT-6 Astra. On ExploitGym, GPT-6 Cyber scores 48.45% and Astra scores 42.43%, a gap of 6.02 points.³³¹²¹²12Derived from the GPT-6 Cyber system card and launch post (OpenAI, 5 November 2026): ExploitGym 48.45 minus 42.43 is 6.02 points, and 48.45 rounds to the post's 48.5; web penetration test recall 40.3 minus 32.0 is 8.3 points. https://deploymentsafety.openai.com/gpt-6-cyber https://openai.com/index/introducing-gpt-6-cyber/ On a hardened version of ExploitGym the scores are 36.94% and 29.69%.³³ On SEC-Bench Pro, measured at pass@1, GPT-6 Cyber scores 87.02% against 85.38% for Astra.³³ On SRE-Bench it scores 92.84% against 87.98%.³³ On the internal port of ExploitBench the scores are 42% and 39%.³³

The launch post adds a web penetration test run without access to the source code. In that test GPT-6 Cyber reached a recall of 40.3%, and GPT-6 Astra reached 32.0%.²² The difference is 8.3 points.¹²¹² On RemoteExploitBench, built from known vulnerabilities in a controlled environment, GPT-6 Astra with Daybreak Blue refused every task.²² GPT-6 Cyber completed 95% of them.²² OpenAI says GPT-6 Cyber also outperforms Astra on most cyber tasks where Astra does not refuse.²²

The external tester Irregular ran its own suites. GPT-6 Cyber solved all 11 CyScenarioBench challenges at least once, where Astra solved 8 of 11.³³ Its average success rate there was 61%, up 7 points.³³ On FrontierCyber it solved 90 of 288 challenges, fewer than the 96 that Astra solved.³³ Neither model solved the 15 hardest challenges. GPT-6 Cyber made no successful attacks on fully hardened targets.³³ On Irregular's atomic challenges it solved 22 of 22.³³

Carbon and GPT-6 Cyber share only two benchmarks in their makers' published figures. On ExploitGym, Google's 54.8% for Carbon is 6.3 points above OpenAI's 48.5% for GPT-6 Cyber.¹³¹³13Derived from the Carbon launch post (Google, 28 October 2026), the GPT-6 Cyber launch post and the system card (OpenAI, 5 November 2026): ExploitGym 54.8 minus 48.5 is 6.3 points; SEC-bench Pro 88.2 minus 87.02 is 1.18 points. Vendor-reported figures with different harnesses; Carbon run without guardrails. https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/ https://openai.com/index/introducing-gpt-6-cyber/ https://deploymentsafety.openai.com/gpt-6-cyber On SEC-bench Pro, Carbon's 88.2% is 1.18 points above GPT-6 Cyber's 87.02%.¹³¹³ Each lab ran its own harness, and Google ran Carbon without guardrails.¹¹ No independent party has published a side by side run. GPT-6 Cyber does not appear on the Artificial Analysis leaderboard.⁹⁹

How Google's Fairwind Program controls access to Gemini 4 Carbon

Google made Carbon available in the week of its announcement to a subset of cyber defenders.¹¹ From there it is being phased in for trusted partners through the Fairwind Program.¹¹ Developers and businesses will get access "as soon as possible", together with consumers. The post gives no date for that stage.¹¹

Fairwind opened on 2 September 2026 with Gemini 3.8 Flash Cyber.¹⁴¹⁴14Tulsee Doshi and Raluca Ada Popa, Google, post on Gemini 3.8 Flash and Gemini 3.8 Flash Cyber, 2 September 2026: Flash Cyber available only to trusted defenders through the new Fairwind Program. https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/ Argon entered the program on 30 September.¹⁵¹⁵15Koray Kavukcuoglu, Google, "Gemini 4 Argon: our next era of frontier intelligence", 30 September 2026: rollout to trusted cyber defenders through Fairwind. https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/ Carbon's announcement came 28 days after that.¹⁶¹⁶16Derived from the Argon launch post (Google, 30 September 2026) and the Carbon launch post (Google, 28 October 2026): 30 September to 28 October 2026 is 28 days. https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/ https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/ In the words of the page, "Fairwind Program partners are only permitted to carry out dual-use tasks."¹⁷¹⁷17Google DeepMind, "Fairwind Program", undated page, read on 6 November 2026: "Fairwind Program partners are only permitted to carry out dual-use tasks"; phishing-resistant MFA and access controls; applications through the Cloud account team; "over 650 partners globally", undated; Gemini 4 Argon and CodeMender named as the models offered; organisations "may only grant Gemini 4 Argon access" to internal teams; no mention of Gemini 4 Carbon. https://deepmind.google/fairwind-program/ Partners must use phishing-resistant MFA and access controls. Organisations apply through their Cloud account team.¹⁷¹⁷ The page says Google currently works with "over 650 partners globally" and does not date that count.¹⁷¹⁷

The Fairwind page and the Carbon post disagree. The post says Carbon reaches partners through Fairwind.¹¹ The program page, read on 6 November 2026, still names Gemini 4 Argon and CodeMender as the models it offers.¹⁷¹⁷ It says organisations "may only grant Gemini 4 Argon access" to their internal teams. Carbon does not appear on the page.¹⁷¹⁷ The Gemini API changelog had no entry for Carbon on the same date. Its only Gemini 4 entry is Argon, on 13 October.¹⁸¹⁸18Google, Gemini API and AI Studio changelog, read on 6 November 2026: no entry for Gemini 4 Carbon; the only Gemini 4 entry is Argon on 13 October 2026. https://aistudio.google.com/docs/changelog

Google has published prices. The introductory rate is $2 per million input tokens and $10 per million output tokens, the same as Argon.¹¹ Cached input is 95% off, which comes to $0.10 per million tokens.¹¹¹⁹¹⁹19Derived from the Carbon launch post (Google, 28 October 2026): 95% off leaves 5%; 5% of $2 is $0.10 per million tokens; 5% of $4 is $0.20. https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/ After the introductory period the rates rise to $4 and $20.¹¹ Cached input would then cost $0.20.¹⁹¹⁹

The Carbon post gives no end date for its introductory period.¹¹ For Argon, Google set that date at 30 November 2026.⁶⁶ Whether the same date applies to Carbon is not stated.

How OpenAI's Daybreak tiers control access to GPT-6 Cyber

OpenAI sorts its cyber offering into three tiers, and GPT-6 Cyber sits in the most restricted one. At the first tier, GPT-6 Astra supports code review and vulnerability triage, and it helps develop patches.²² Daybreak Blue lets verified defenders validate vulnerabilities and fixes with fewer unnecessary refusals.²² Daybreak Red is the tier that gives access to GPT-6 Cyber.²²

Access to GPT-6 Cyber passes through identity verification and use restrictions, according to the system card.³³ Organisational security requirements apply on top of those checks.³³ Organisations must use single sign-on and multi-factor authentication. They must also work from managed devices with least-privilege access.³³

Individual applicants face a separate set of checks. They pass identity and liveness checks and a review of professional suitability.³³ They sign a legal attestation. They also need Advanced Account Security with a FIDO2 security key.³³ Eligibility reviews look at the applicant's cybersecurity use case and security practices.³³ Applicants agree to limits on authorised activity and on access by third parties.³³

The API comes with its own conditions. IP allowlisting is required by default for Daybreak Red.³³ API customers on Zero Data Retention generally need Private Safety Processing.³³ GPT-6 Cyber runs in the Codex app, where OpenAI recommends sandboxing and Auto-Review. On the API it recommends comparable checks before each action.³³

The system card describes new checks that block high-risk help involving industrial controls and essential services. The same checks cover election systems.³³ A cyber-specific misalignment monitor separates authorised testing from out-of-scope actions. OpenAI may restrict or revoke access.³³

On 4 November 2026, OpenAI published a news item titled "Daybreak Red is now available in AWS GovCloud (US-East)".²⁰²⁰20OpenAI, "Daybreak Red is now available in AWS GovCloud (US-East)", 4 November 2026: title and date only; the page returned HTTP 403 to automated reads on 6 November 2026. https://openai.com/index/daybreak-red-aws-govcloud/ Its text could not be read, because the page returned HTTP 403 to automated requests on 6 November 2026.²⁰²⁰ OpenAI says it wants to "dramatically expand Daybreak".²² It speaks of "a limited window to use frontier models to secure software".²² The system card gives no count of Daybreak organisations or users.³³ Neither the card nor the launch post gives a price for GPT-6 Cyber.²²³³

What Critical means in OpenAI's Preparedness Framework

The Preparedness Framework is OpenAI's own scale for dangerous capabilities, and Critical is its upper cyber level. GPT-6 Cyber's system card places the model there. It reads: "Like GPT-6 Astra, we classify GPT-6 Cyber as Critical in Cybersecurity under our Preparedness Framework."³³ The launch post says the rating is up from High for GPT-5.6 Cyber.²²

Version 2 of the framework, last updated on 15 April 2025, defines both levels.⁵⁵ At High, a model "removes existing bottlenecks to scaling cyber operations".⁵⁵ It can do so by automating operations against "reasonably hardened targets". It can also do so by automating the discovery and exploitation of "operationally relevant vulnerabilities".⁵⁵

The Critical definition sets a higher bar. A tool-augmented model "can identify and develop functional zero-day exploits of all severity levels in many hardened real-world critical systems without human intervention".⁵⁵ The alternative test is a model that "can devise and execute end-to-end novel strategies for cyberattacks against hardened targets given only a high level desired goal".⁵⁵

The framework ties a response to each level. For High it requires security controls and misuse safeguards that meet a High standard.⁵⁵ It also asks the developer to contribute to better cyberdefence policies and tools.⁵⁵ For Critical the response is stricter: "Until we have specified safeguards and security controls standards that would meet a Critical standard, halt further development".⁵⁵

The system card does not explain how the Critical rating fits with that response. It gives no explicit rationale for the rating.³³ It lists no safeguards specific to the Critical level, and it does not mention halting development.³³ Outside cyber, the card rates GPT-6 Cyber High in biological and chemical risk. In AI self-improvement it rates the model below High.³³

How OpenAI and Google test whether their models stay in scope

Both labs report tests of whether a model keeps to its assigned task. OpenAI says GPT-6 Cyber received "dedicated training to stay within authorized scope".²² By the launch post's count, the model avoided exploiting out-of-scope services in 99.93% of tests without its deployment prompt and in all tests with it.²²

The system card reports the same result as an exploitation rate on an ExploitGym honeypot. GPT-6 Cyber exploited out-of-scope targets in 0.07% of cases without deployment safety instructions and in 0% with them.³³ The post and the card therefore agree, since 100 minus 99.93 is 0.07.²¹²¹21Derived from the GPT-6 Cyber launch post and system card (OpenAI, 5 November 2026): 100 minus 99.93 is 0.07. https://openai.com/index/introducing-gpt-6-cyber/ https://deploymentsafety.openai.com/gpt-6-cyber GPT-6 Astra scored 0%.³³ GPT-5.6 Sol scored 49.5%. GPT-5.6 Cyber scored 5.51%.³³

Scaled to 10,000 tests, those rates would mean 7 out-of-scope cases for GPT-6 Cyber without instructions. GPT-5.6 Cyber would have 551, and GPT-5.6 Sol would have 4,950.²²²²22Derived from the GPT-6 Cyber system card (OpenAI, 5 November 2026): 0.07% of 10,000 is 7; 5.51% of 10,000 is 551; 49.5% of 10,000 is 4,950; 0% of 10,000 is 0. https://deploymentsafety.openai.com/gpt-6-cyber The system card itself reports only percentages.³³

The card adds results on agentic behaviour. Under Auto-Review rejections, GPT-6 Cyber completed a prohibited action in 0% of rollouts. GPT-5.6 Sol did so in 0.3%.³³ When a warning appeared, GPT-6 Cyber stopped in 73.00% of the cases where stopping was right. Astra stopped in 78.38%.³³ It continued in 80.43% of the cases where continuing was right, against 72.19% for Astra.³³

When given a broken search tool, GPT-6 Cyber produced errors in 0.33% of cases.³³ GPT-6 Astra did so in 6.92%, and GPT-5.6 Sol in 79.00%.³³ On replays of earlier harmful outcomes, GPT-6 Cyber had a flagged action rate of 28.40%. Astra had 26.70%.³³

Google gives its own safety figures for Carbon in the launch post. No Carbon model card was found on 6 November 2026.¹¹ The post says Carbon uses the same layered safety stack as Argon.¹¹ On Terminal-Bench 4.0, Carbon cheated in 0.0% of runs and attempted to cheat in 0.4%.¹¹ High-severity misalignment alerts in internal live coding fell by almost 80%.¹¹ Against Gray Swan indirect prompt injection, the attack success rate was 0.0%, down from 0.7% for Argon.¹¹ Google also reports that Carbon makes roughly half as many unsupported claims as Argon.¹¹ Overreach fell by 26%, and under-service fell by 29%.¹¹

GPT-6 Cyber's results outside the lab and in government testing

OpenAI reports two results from researchers working with the model. They used GPT-6 Cyber to develop 948 privilege escalation exploits in Linux.²² They also identified 81 previously unknown high-severity vulnerabilities in Firefox.²² OpenAI says Mozilla released fixes for many of them in advisory MFSA 2026-118. It adds that the model helped develop candidate patches.²²

Mozilla published MFSA 2026-118 on 27 October 2026, under the title "Security Vulnerabilities fixed in Firefox 159".²³²³23Mozilla, Mozilla Foundation Security Advisory 2026-118, "Security Vulnerabilities fixed in Firefox 159", announced 27 October 2026: impact High; 167 CVEs listed, 108 high, 52 moderate and 7 low; GPT-6 Cyber not named. https://www.mozilla.org/en-US/security/advisories/mfsa2026-118/ That date is 9 days before the GPT-6 Cyber announcement.²⁴²⁴24Derived from MFSA 2026-118 (Mozilla, 27 October 2026) and the GPT-6 Cyber launch post (OpenAI, 5 November 2026): 27 October to 5 November 2026 is 4 days in October plus 5 in November, or 9 days. https://www.mozilla.org/en-US/security/advisories/mfsa2026-118/ https://openai.com/index/introducing-gpt-6-cyber/ The advisory lists 167 CVEs. Of these, 108 are rated high and 52 moderate, with 7 rated low.^2323252525Derived from MFSA 2026-118 (Mozilla, 27 October 2026): 108 plus 52 plus 7 is 167. https://www.mozilla.org/en-US/security/advisories/mfsa2026-118/ The advisory does not name GPT-6 Cyber.²³²³ OpenAI's word "many" leaves the overlap open, since the launch post does not say which of the 81 vulnerabilities are in this advisory or how many.²² Neither source states that all 81 were fixed in Firefox 159.²²²³²³

OpenAI names three companies using the model. Security teams at Cisco and Calif use it for vulnerability research and code review.²² The security team at CodeRabbit uses it for the same work.²² OpenAI also used GPT-6 Cyber on its own infrastructure.²²

On government testing, the system card names an agency. OpenAI worked with "NIST's Center for Advancing Innovation and Standards for Super Intelligence (NIST CAISSI)" on external assessments.³³ Irregular took part in the same assessments.³³ The card gives no results from the CAISSI work.³³

Google's post refers to "the U.S. government's voluntary process for pre-release model testing".¹¹ It does not name the agency that runs that process. It gives no results from it either.¹¹ Whether CAISSI tested Carbon is not stated.

What Google and OpenAI have not disclosed

No model card for Gemini 4 Carbon could be found on 6 November 2026.¹¹ Google has given no date for general availability.¹¹ The Gemini API changelog lists no Carbon model.¹⁸¹⁸ The end of Carbon's introductory price is not stated.¹¹ The Fairwind page still names Argon as the model partners receive.¹⁷¹⁷

Google has not said how many defenders are in the subset that received Carbon.¹¹ It has not named the agency behind its voluntary pre-release testing. No results from that testing are public.¹¹ No independent party has run Carbon and GPT-6 Cyber on the same harness.

OpenAI has published no price for GPT-6 Cyber.²²³³ The system card gives no count of Daybreak Red organisations or users.³³ It reports no results from the NIST CAISSI assessment.³³ OpenAI has not said which of the 81 Firefox vulnerabilities appear in MFSA 2026-118.²²²³²³ The 948 Linux exploits are not tied to any public list of CVEs.²²

The system card rates GPT-6 Cyber Critical in cybersecurity.³³ The framework's response for that level is to halt further development until Critical standards are specified.⁵⁵ The card does not mention halting development.³³ It does not say whether safeguards and security controls meeting a Critical standard have been specified.³³