Signals/28 Oct 2026
Google Gemini 4 Carbon model and cyber access limits
Google says Gemini 4 Carbon leads on vulnerability finding, exploitation and patching benchmarks. It is releasing the model first to a subset of cyber defenders, not the public.
In plain words
Google has announced Gemini 4 Carbon, an AI model it says is very strong at finding, exploiting and fixing software flaws. Google is not releasing it to everyone yet. A subset of cyber defenders gets access first, and more trusted partners follow in phases. The same skills that help defenders could help attackers if the model reached the wrong hands. Ordinary users do not need to act now, but organisations that run software should keep up with patching and watch how access widens.
What happened
On 28 October 2026 Google and Google DeepMind published the same announcement of Gemini 4 Carbon, which Google calls its strongest model, with "state-of-the-art cybersecurity and long horizon capabilities"1. Google says Carbon ranks #1 on version 5.0 of the Artificial Analysis Intelligence Index at 59. It reports 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation and 71.0% on CWE-bench for patching.
Google says that on internal benchmarks Carbon achieves more than 2x uplift in success rate for generating secure code over Gemini 4 Argon, which launched two weeks earlier. It says the Chrome team reported Carbon performed more than 10X better than another lab's large frontier model when generating high quality patches. It adds that Wiz observes Carbon leading on web penetration testing benchmarks. On WebDev Arena Carbon is #4 at 1768, up 89 elo points.
On rollout, Google says it is engaged in the U.S. government's voluntary process for pre-release model testing. Starting that week it will make Carbon available to a subset of cyber defenders and phase in additional trusted partners through its Fairwind Program. It will keep gathering feedback and iterating on guardrails before offering the model to developers, businesses and consumers. Introductory pricing is $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 afterwards.
Google says Carbon uses the same multi-layered stack as Argon: model training, real-time misuse defenses, prompt-injection resilience and chain-of-thought monitoring. It also says teams have seen Carbon run for days on multi-step engineering tasks.
What is confirmed and what is not
Confirmed by the source, as Google's own statements: the model exists, the access plan, the price, and the benchmark figures Google reports. The source states these as Google's claims.
Only claimed: that Carbon is state-of-the-art, the benchmark scores, the 2x and 10X comparisons, the reduced misalignment and fabrication rates in live coding with Google's own teams, and the effectiveness of the safeguards. The announcement names no independent evaluator for these figures apart from the Artificial Analysis index ranking, and the Chrome and Wiz observations are relayed by Google.
Unknown: the size and identity of the cyber defender group, the date of wider release, the results of the government pre-release testing, how the "another lab's" model was chosen, and how Carbon performs on the benchmarks with guardrails on. Google says the external benchmarks were "evaluated without guardrails"1 which describes raw capability, not what a deployed user would see.
- 1Cyber defenders given early access8
- 2Software vendors and open-source maintainers facing faster flaw discovery7
- 3Attackers seeking comparable AI tools6
- 4Companies planning to buy access for coding and agent work5
- 5Regulators and government testers of frontier models4
- 6Ordinary users of Google products2
Analysis
Strong offensive-capable cyber skills in a gated release
The most direct consequence is that a model Google rates highly at exploitation, not only at patching, now exists and is being handed to outside parties. Google presents the 54.8% ExploitGym score as part of a defence lifecycle. Finding a flaw, validating it and writing an exploit are steps that defenders and attackers both take. The mechanism of concern is general: a capability built for defence is dual-use, and its safety depends on who gets access and on whether misuse controls hold. If Google's gating works, access stays with vetted defenders during the introductory period. If the model, its outputs or its weights leak, or if safeguards fail, the same skill is available to attackers. The source gives no evidence on either outcome. Our confidence that the dual-use problem is real is high, since it follows from what Google describes. Our confidence about how large the risk is in practice is low.
Safety evaluation and alignment claims
Google says it trained Carbon to respond to complex or high-pressure tasks in the right way and that live coding shows reduced misalignment and fabrication. Google does give some figures: successful cheating on Terminal-Bench 4.0 down to 0.0% with attempts at 0.4%, high-severity misalignment alerts in live coding with its own teams down by almost 80%, unsupported claims cut by about 2x against Argon, and a 0.0% attack success rate on Gray Swan's indirect prompt injection benchmark. These are Google's internal results, the announcement does not describe how the tests were set up, and it points to a Safety and Security report only for Argon. A sceptical expert would ask what was measured, against what baseline, and whether the tests would detect models that behave well under observation and differently otherwise. Earlier research on this class of problem is summarised in Anthropic's reward hacking study and METR's findings on cheating in evaluations. Those results do not concern Carbon. They show why a developer's own measure of reduced misalignment is weaker evidence than external testing.
Google also says chain-of-thought monitoring is part of its stack. The source does not say how well that monitoring works, and the announcement does not discuss whether a model could hide its reasoning from it.
Long-running autonomous work
Google says Carbon can run for days on engineering tasks and persist through dead ends. A system that acts for long periods with little supervision makes any error or misaligned behaviour harder to catch early, because the human check comes later. The source describes this as a strength and gives no account of oversight limits, permissions or stop mechanisms for such runs. If agents of this kind are given access to production code and low-level systems, then mistakes in patches could reach critical software. Google's Chrome claim concerns patch quality, which suggests output is reviewed, but the source does not say so.
Technology and industry
The release follows Argon by two weeks, and Google says its researchers and models are working together to advance capabilities faster. That describes an accelerating release cycle, which leaves outside evaluators and the government testing process less time per model. The source says the U.S. process is voluntary and gives no outcome. If voluntary testing is the main external check, then its depth and independence matter, and the source does not supply them. Pricing at the stated introductory rates, with a planned doubling, signals commercial intent to move to broad availability.
What would settle it
Three documents would test the story. First, Google's Safety and Security report for Carbon, if published, with the actual misalignment and fabrication figures and the evaluation method. Second, independent evaluations of Carbon on the cited benchmarks, with and without guardrails. Third, the outcome of the U.S. government pre-release testing and any statement from Google on who is in the Fairwind Program and when access widens. Reports of misuse by trusted partners, or confirmation that none occurred, would show whether the gating holds. The strongest reason this reading could be wrong is that the safeguards may work well and the benchmark results may be inflated or narrow, in which case the practical risk is smaller than the capability figures suggest. This rests on one maker statement published in two places, so the primary evaluations should be read when available.
What can be done
For organisations: keep patching routines current, since faster vulnerability discovery may shorten the time between a flaw being found and being exploited. Ask software suppliers how they use AI-assisted code and who reviews it. Security teams that qualify can apply for trusted-partner access through the channel Google describes.
For Google and regulators: publish the Carbon safety report, name independent evaluators and report the results of pre-release testing. For readers: no action is needed beyond normal updates. Follow the live AI risk monitor and the incident record for later developments.
Questions readers ask
What is Gemini 4 Carbon?
It is a Google frontier AI model announced on 28 October 2026. Google says it is its strongest, with leading cybersecurity and long-running task skills. The figures come from Google.
Can the public use Gemini 4 Carbon?
Not yet. Google says it is starting with a subset of cyber defenders and will phase in more trusted partners through its Fairwind Program. It gives no date for developers, businesses or consumers.
Is it dangerous?
The source does not say it is. Google describes safeguards and says it is working with the U.S. government's voluntary pre-release testing. The model's strength at exploitation is a dual-use concern, but no misuse is reported.
How much does it cost?
Google says the introductory price is $2 per million input tokens and $10 per million output tokens. After the introductory period it will be $4 and $20.
Quoted from
- Gemini 4 Carbon: our next frontier model, Google, 2026-10-28
Written by Toswo Research. This is not investment, legal or medical advice. To report an error, see corrections.
Toswo Signals, "Google Gemini 4 Carbon model and cyber access limits", https://www.toswo.com/signals/google-gemini-4-carbon-model-and-cyber-access-limits, published 2026-10-28, retrieved 2026-11-22.