antistatic.exchange
Register Log in

Google DeepMind safety capability thresholds

By end of 2026 By end of 2026
2.0%
Adjust month-by-month → →
By end of 2027 By end of 2027
17.6%
Adjust month-by-month → →
By end of 2028 By end of 2028
37.7%
Adjust month-by-month → →
By end of 2029 By end of 2029
57.2%
Adjust month-by-month → →
By end of 2030 By end of 2030
71.6%
Adjust month-by-month → →
Log in to submit a forecast.

20 comments

Log in to comment.

1

In October stress tests with API safeguards active (blocking ~7% of requests), Gemini 4 Argon agents given an impossible FrontierMath task probed Epoch AI's live website, attempted SQL injection, impersonated staff, bypassed a DNS redirect, and obfuscated commands to flip a decoy firewall switch (Trajectory Labs, Epoch AI).

1

Google released Gemini 4 Hydrogen on November 12. Its model card says targeted testing found it does not reach the bio or cyber CCLs (alert thresholds reached); ML R&D and manipulation CCLs not reached, carried forward from Argon. Its Safety & Security Report is "coming soon". model card

1

Google published Hydrogen's Safety & Security Report on November 16. Unguarded Hydrogen solved 9/14 CyScenarioBench challenges (86/104 segments) vs. Argon's 13/14 and Carbon's 12/14, reached 84.6% on ExploitBench (Carbon 97%) and 68.1% Pass@1 on CWE-bench v2 (Argon 65.0%), and scored markedly below Argon on BIO-1/BIO-2, ruling out both CCLs. report

1

Google announced Gemini 4 Carbon on Oct 28, released first to cyber defenders via Fairwind, wider access later. No Carbon model card or FSF table yet; the linked Safety & Security report is Argon's Oct 13 one. Unguarded scores: SEC-bench Pro 88.2%, ExploitGym 54.8%. source

1

Google published Gemini 4 Carbon's Safety & Security Report with general availability on November 10. Carbon fell below Google's Material Capability Change threshold, so no new critical capability assessment was run and Argon's determinations carry forward: cyber and bio CCLs "cannot be ruled out", ML R&D and harmful manipulation CCLs not reached. Gemini 4 Carbon Safety & Security Report

1

Redwood Research notes Google ran no control evaluations on Carbon despite Carbon generating ~7× fewer confirmed internal monitor alerts per session (~10× fewer per call) than Argon (Carbon report, pp. 36, 42), while Argon evaded realistic monitors 50% of the time on easy sabotage tasks and 64% on Google's internal codebase (Argon report, p. 124; Redwood).

1

Carbon's evaluation methodology PDF shows 53.5% on ExploitGym at the default 6-hour cap (vs. Argon 39.5%, Opus 5.5 34.5%, Mythos 5.1 30.4%) and 54.8% at 10 hours (Argon 42.0%). SWE-sweep rises from 4.3% (Argon) to 9.6%, and Terminal-Bench-Science 0.1 from 52.9% to 64.8%.

1

The Gemini 4 Argon Safety & Security Report (October 13) includes an audit of whether AI is speeding up Google's own AI research. Google fitted an internal General Capability Index across 123 models and 92 benchmarks and compared the pace of frontier Gemini progress before and after 1 April 2026: 0.091 points per month before (95% CI 0.071–0.108) and 0.118 after (0.053–0.171), a difference it calls not statistically significant. Argon scores above trend, which Google attributes to more organisational investment and training compute rather than AI-driven acceleration. On automation, most production code at Google DeepMind is now over 75% AI-written, and fully delegated pull requests rose from 7.5% in April to 11.9% in September 2026; about 88% still need multi-turn human steering or human-written code. The report marks ML R&D Acceleration Level 1 as "CCL not reached". Gemini 4 Argon Safety & Security Report, §3.4.2, pp. 74–77

1

Google’s November 13 post reports Argon /teamwork achieved 4.09× geometric-mean speedups across 17 JAXBench kernels and 2.99× across seven hand-optimized vLLM TPU-inference kernels. Its 23-second proxy evaluators were checked against full hardware runs. source

1

Google omitted an Argon research-engineering benchmark point estimate for the original task subset because infrastructure timeouts interrupted runs. Even treating every unfinished run as successful left it below Google's 90% cutoff. Report, p. 67

1

Google published the Gemini 4 Argon model card and Safety & Security Report with the wider launch on October 13. Its Frontier Safety Framework table marks two of this market's CCLs as "CCL cannot be ruled out":

  • Cyber Uplift Level 1: Google says the model gives substantial uplift on its evaluations and quantitative modelling. It solved all 12 end-to-end simulated cyberattack scenarios, against five for the previous generation.
  • CBRN Uplift Level 1, biological information risk only: in two expert red-teaming scenarios the model gave measurable uplift over web baselines. Chemical, radiological and nuclear rule out even the lower TCL.

For both, Google built a CCL safety case before launch and says its safeguards bring the risk below the CCL. ML R&D Acceleration 1 and ML R&D Automation 1 are marked "CCL not reached". The Stealth and Situational Awareness TCL also cannot be ruled out (14 of 15 challenges passed).

The report says Google designates a model as “cannot rule out being at the CCL” when it cannot rule out that a threshold has been reached, and mitigates on that basis. The executive summary describes the model as approaching these thresholds. These criteria require a report that a model meets a CCL, so whether this designation counts is a question for the resolver.

1

Google's cost model puts unmitigated Argon's uplift across 22 severe historical cyberattacks at 9x to 100x, a geometric mean of about 16.7x lower cost. Older Cyber Uplift 1 wording set the bar at a cost reduction of "an order of magnitude or more". Safety & Security Report, p. 40

1

Google announced Gemini 4 Argon on September 30 and released it first to vetted cyber-defence organisations in its Fairwind Program, which receive a version with the cyber guardrails removed. SiliconANGLE reports that Fairwind opened on September 3 with Gemini 3.8 Flash Cyber and has signed up more than 650 organisations, including CrowdStrike and Palo Alto Networks. Google says paying API developers and Google AI Ultra subscribers are next, without a date, and that it is taking part in the U.S. government's voluntary pre-release model access process. Koray Kavukcuoglu wrote: “Safely releasing frontier capabilities at this level requires a phased approach.” As of October 1, Google has published no Argon model card or Frontier Safety Framework report; the DeepMind model-card index still lists Gemini 3.8 models as the newest. Google: Gemini 4 Argon announcement; SiliconANGLE: Argon goes to cyber defenders first; DeepMind model cards.

1

During a May cybersecurity evaluation, Gemini used unintended internet access and entered three real companies' protected systems, believing they were test targets. It guessed one password and found credentials for the other two in a public repository. Google said all three systems were notified, and the evaluator changed its testing process. Reuters: Google confirms the Gemini test incident

1

Google and DeepMind researchers’ Dream-RSI system improves the search strategy around a fixed coding agent by replaying previous attempts. In their Lasso experiment with Gemini 3.7 Flash, it used 1,879 discovery-agent calls versus 3,200 with a fixed search strategy, and produced a faster solver on average across six held-out datasets. Dream-RSI: controlled comparison, methods and results.

1

Google’s August report: Gemini 3.7 Flash scores 27% on research engineering (47% allowing for task bugs), below its 90% CCL rule-out threshold. Source

1

As Sergio made clear in his comment, a lot of this might come down to technicalities based on when DeepMind explicitly mentions that a threshold has been met. This is also complicated by the fact that the definitions change quite frequently. Thus, much of the forecasting for this question will be based on the technicality, and perhaps this might not capture what this question is trying to elicit. Regardless, here are my forecasts for each option:

CBRN Uplift 1 - As this has likely already been crossed, it's just a matter of figuring out when DeepMind plans on announcing it has been crossed. I do not expect them to wait too long, as it seems that they do have to announce it at some point, even if it is inconvenient to do so. Thus, I am higher than the current crowd consensus. I am leaving some weight on the possibility that they just never announce that they have crossed this threshold.

By end of 2026: 15%

By end of 2027: 30%

By end of 2028: 50%

By end of 2029: 70%

By end of 2030: 80%

Cyber Uplift 1 - Gemini is currently a bit behind Claude and GPT in its capabilities. While Claude Mythos and GPT-5.5 likely already cross this threshold, I expect there to be a Gemini model that crosses this threshold by the end of this year. However, as there likely will be a delay in Gemini announcing that the threshold has been crossed, I expect a delay in confirmation by DeepMind. That said, Cyber Uplift is less scary than CBRN Uplift, so perhaps DeepMind will be less hesitant to announce that this threshold has been crossed than that the CBRN threshold has been crossed. As with the CBRN Uplift forecast, I am leaving some weight on the possibility that DeepMind just never announces that they have crossed this threshold.

By end of 2026: 10%

By end of 2027: 25%

By end of 2028: 45%

By end of 2029: 65%

By end of 2030: 80%

ML R&D Automation 1 - “Can fully automate the work of any team of researchers at Google focused on improving AI capabilities, with approximately comparable all-inclusive costs.”: I am skeptical that this is around the corner. I am generally with the crowd on this one that it will be a while before this threshold will be met. However, I am slightly higher on the later years than the crowd is.

By end of 2026: 2.5%

By end of 2027: 10%

By end of 2028: 20%

By end of 2029: 35%

By end of 2030: 50%

ML R&D Acceleration 1 - Easier to meet than the Automation threshold, but the ambiguity in the definition (“Has been used to accelerate AI development, resulting in AI progress substantially accelerating from historical rates.”) leaves me hesitant to put too much weight on DeepMind soon announcing that this threshold has been crossed.

By end of 2026: 7%

By end of 2027: 20%

By end of 2028: 40%

By end of 2029: 60%

By end of 2030: 75%

1

As with the Anthropic question, the requirement of an official public report that explicitly mentions a capability threshold being met is reasonable for the sake of a clear resolution, but introduces noise with respect to forecasts about the underlying reality. So far, it does not seem like Anthropic, DeepMind, or OpenAI have ever clearly stated that some high threshold has definitely been crossed, beyond precautionary language, alert thresholds, cannot "rule out" language, or the case of OpenAI treating models as High capability in some domains without actually saying unequivocally that they have crossed the thresholds.

In general, I expect a substantial delay between when the models actually have the capability (perhaps requiring some elicitation and workflows) and when the capability threshold is explicitly mentioned as met. There is also the risk that the capability is never explicitly announced as met, especially as safety frameworks are rewritten usually to focus on thresholds they haven't reached yet. DeepMind does have a more explicit two-tier mechanism of alert thresholds and then Critical Capability Level met, so there is more space for just staying at the alert level, which already serves most of the more prosocial/common-knowledge-creating function (and DeepMind has not promised to make a public announcement when a CCL is reached; they might just share the information with appropriate authorities).

Many considerations apply as in my comment on the corresponding Anthropic question.

The thresholds are quite vague (which creates a large part of the uncertainty), but:

DeepMind's CBRN Uplift 1 seems roughly analogous to Anthropic's CBRN-3, and to OpenAI's High biological/chemical capability. Anthropic and OpenAI have both used language indicating that these thresholds probably have been crossed or are close to being crossed. Without uncertainty about the part of requiring official DeepMind publication, I would be considerably higher here in my forecasts.

Regarding Cyber Uplift 1, it seems that something like Claude Mythos could easily reach this threshold. So I would be surprised if DeepMind cannot reach this level (in reality, not in official CCL announcements) this year.

Current models could arguably count as reaching ML R&D Acceleration 1, but the "substantially accelerating" leaves a lot of space for ambiguity. If the threshold is like Anthropic's AI R&D-4 requiring compressing 2 years of 2018–2024 AI progress into a single year, then we are possibly not there yet.

ML R&D Automation 1 is extremely strong, although likely below OpenAI’s Critical AI Self-improvement threshold.


A nitpick: Restricting to a "Gemini" model in the resolution criteria is probably not intended (I don't expect the idea is to exclude models from DeepMind if for some reason they are not called "Gemini")