Google DeepMind says its Gemini 4 Carbon safety assessment included tests specific to Carbon, but it carried forward some capability determinations from Gemini 4 Argon after Carbon fell below Google’s material-change threshold. In an analysis published November 16, 2026, Redwood Research argues that the reports do not demonstrate low loss-of-control risk because dedicated control evaluations were not repeated on Carbon. Google’s November 2026 safety report sets out the company’s assessment.

A cybersecurity benchmark, a capability threshold, and a control evaluation answer different questions. Keep their results in separate lanes.

Why did Google carry forward some Argon findings?

Google’s Material Capability Change assessment, or MCC, uses GCI v2, an internal index that aggregates benchmark performance. Google considers a change material when it exceeds four months of capability progress; the report estimates historical progress at roughly 4.7 GCI points per month. Google says Carbon fell below that threshold, so it carried forward Argon’s capability evaluation results and threshold determinations instead of conducting a full new critical-capability assessment.

Among those carried-forward determinations, Google says it could not rule out the Cyber Uplift Level 1 and biological-information CBRN Uplift Level 1 critical capability levels, or the Stealth and Situational Awareness tracked capability level. CBRN refers to chemical, biological, radiological, and nuclear risks.

For earlier context on a separate debate about Argon’s practical coding performance, NeoTeo covered questions surrounding Gemini 4 Argon’s coding capabilities.

What did Google assess specifically on Carbon?

Google lists several Carbon-specific evaluations: baseline CBRN propensity; coverage and adversarial robustness of CBRN and cybersecurity safeguards; a condensed alignment assessment; and checks of chain-of-thought transparency, including legibility, coverage, and monitorability.

The report’s evaluations generally measured the model without production safeguards operating outside the model and without search grounding, unless a test was specifically described otherwise. Google separately describes deployment-time protections, including prompted classifiers, activation-based introspection probes, pattern and embedding classifiers, and chain-of-thought monitors. The protections can vary by deployment surface.

Google also reports a 54.8% causal pass rate for Carbon on ExploitGym, a cybersecurity benchmark. That result concerns performance on cybersecurity tasks; control evaluations address whether a model can evade monitors.

What does Redwood Research’s criticism say?

Redwood Research says Google did not repeat dedicated control evaluations on Carbon, including tests involving monitor evasion, steganography, and chain-of-thought controllability. The evasion rates Redwood cites—about 45.5% in one LinuxArena configuration and about 64% in a separate internal-codebase setting—were results for Argon, not Carbon. The two figures come from different settings.

Google also reports fewer confirmed high-severity monitor alerts for Carbon in internal use: about seven times fewer per session and ten times fewer per model call. Google notes that the comparison involved differences in tasks, trajectories, review coverage, and checkpoint selection. Redwood argues that the alert counts alone cannot distinguish less misbehavior from greater difficulty monitoring Carbon.

Redwood’s authors, Adam Kaufman and Alek Westover, say they support external deployment while calling for stronger safety arguments, especially for models used internally in AI research and development.