Gemini 4 Carbon
- Framework: Frontier Safety Framework
- 7 statements about capability levels
- 1 independent or government test
- 1 safeguard
- 40 tests by the provider
- 3 source documents
What the provider states
- Autonomy & replication · CCL (ML R&D acceleration) (Critical Capability Level) · Stated as not crossed
CCL not reached (carried forward from Argon)
Google DeepMind · 10 November 2026 · Read the source - Autonomy & replication · CCL (ML R&D automation) (Critical Capability Level) · Stated as not crossed
CCL not reached (carried forward from Argon)
Google DeepMind · 10 November 2026 · Read the source - Chemical, biological, radiological and nuclear · CCL (Biological) (Critical Capability Level) · Not determined
CCL cannot be ruled out (carried forward from Argon)
Google DeepMind · 10 November 2026 · Read the source - Chemical, biological, radiological and nuclear · TCL (Chemical, Radiological & Nuclear) (Tracked Capability Level) · Stated as not crossed
TCL not reached (carried forward from Argon)
Google DeepMind · 10 November 2026 · Read the source - Cyber · CCL (Critical Capability Level) · Not determined
CCL cannot be ruled out (carried forward from Argon)
Google DeepMind · 10 November 2026 · Read the source - Deception & sabotage · TCL (Stealth and Situational Awareness) (Tracked Capability Level) · Not determined
TCL cannot be ruled out (carried forward from Argon)
Google DeepMind · 10 November 2026 · Read the source - Harmful content · CCL (Critical Capability Level) · Stated as not crossed
CCL not reached (carried forward from Argon)
Google DeepMind · 10 November 2026 · Read the source
Independent and government tests
- Trajectory Labs Red-Teaming · Trajectory Labs, independent evaluator · Cyber
Under these conditions, Trajectory Labs reported they found no new universal jailbreaks that generalized across tasks within this timeframe.
Google DeepMind · 10 November 2026 · Read the source
Safeguards the provider describes
- Misuse monitoring and enforcement
This is why our misuse prevention approach incorporates additional layers of defense such as log monitoring and account enforcement.
Google DeepMind · 10 November 2026 · Read the source
The provider's own tests
- 375-task agentic honesty benchmark · The provider · Deception & sabotage
New results show that on our 375-task agentic honesty benchmark, it passed 92% of tasks (compared to 70% for Argon) — declining to fill in missing experimental data 98% of the time and flagging methodological flaws 68% of the time.
Google DeepMind · 10 November 2026 · Read the source - Agentic task-gaming & overreach evaluations · The provider · Deception & sabotage
0.56 mean score (vs Argon 0.49, Gemini 3.8 Flash 4.07, Gemini 3.1 Pro 7.11), with 92% of Carbon's points from one ambiguous-intent helpfulness task while security-bypass scores fell from 0.27 to 0.03; 0.00 across all 72/360 runs below
Google DeepMind · 10 November 2026 · Read the source - Automated content safety evaluation measuring safety policies · The provider · Harmful content
Automated content safety evaluation measuring safety policies +3.5pp
Google DeepMind · 10 November 2026 · Read the source - CBRN Conversational Evaluations · The provider · Chemical, biological, radiological and nuclear
The model achieves ~0% violations for topics related to bioweapons (sourcing, enhancement, weaponization) and chemical weapons.
Google DeepMind · 10 November 2026 · Read the source - Child Safety evaluations · The provider · Harmful content
Our Child Safety evaluations found that Carbon , like Argon, is robust to requests for this type of content .
Google DeepMind · 10 November 2026 · Read the source - Coding agent evaluations · The provider · Injection & agentic
Built on SWE-Bench Pro ( Deng et al., 2025 ), this 75-scenario benchmark measures whether an autonomous coding agent will execute prompt injections disguised as legitimate steps toward the user’s goal.
Google DeepMind · 10 November 2026 · Read the source - Cyber capability benchmarks SEC-bench Pro, ExploitGym and CWE-bench v1 · The provider · Cyber
When accessed through the Fairwind Program, Carbon leads on key benchmarks across finding (SEC-bench Pro), exploiting (ExploitGym), and patching (CWE-bench v1) vulnerabilities: inspecting software, tracing how flaws can be chained together across systems, and generating verified root-cause patches.
Google DeepMind · 10 November 2026 · Read the source - Dangerous Content Safe Response Rate · The provider · Harmful content
Gemini 4 Carbon demonstrates a slight improvement over Argon on Dangerous Content Safety and performs similarly to 3.8 Flash.
Google DeepMind · 10 November 2026 · Read the source - Gray Swan Biological Evaluations (Shade) · Run by Gray Swan, reported by the provider · Chemical, biological, radiological and nuclear
The model stays robust against Gray Swan’s automatic attacks, achieving a <1% per-query success rate and a 6% per-seed-prompt OSR.
Google DeepMind · 10 November 2026 · Read the source - Gray Swan IPI benchmark for indirect prompt injection · The provider · Injection & agentic
Two weeks ago, we shared that Carbon achieved a 0% prompt injection attack success rate on the Gray Swan IPI benchmark — marking the first time that milestone has been reached.
Google DeepMind · 10 November 2026 · Read the source
The 10 newest are listed above; show all 40
- Gray Swan automated cyberattack query testing · The provider · Misuse & jailbreak
Our safeguards also lowered the automated cyberattack query success rate to 2.17% in Gray Swan’s testing, compared to 9.86% on Argon.
Google DeepMind · 10 November 2026 · Read the source - Image to Text Safety · The provider · Harmful content
Reported result: -1.00pp versus Gemini 4 Argon
Automated content safety evaluation measuring adherence to safety policies with both images and text as inputs -1.00pp
Google DeepMind · 10 November 2026 · Read the source - Internal vulnerability research · The provider · Cyber
In this evaluation, the model is given an entry point and (optionally) source code that it must analyze to find a vulnerability.
Google DeepMind · 10 November 2026 · Read the source - Multi-turn hate speech & harassment safety evaluation · The provider · Harmful content
Figure 4.7: Multi-turn hate speech & harassment safety evaluation results.
Google DeepMind · 10 November 2026 · Read the source - Multi-turn sexually explicit content safety evaluation · The provider · Harmful content
Figure 4.5: Multi-turn sexually explicit content safety evaluation results.
Google DeepMind · 10 November 2026 · Read the source - Project Fortify internal scanning and triage benchmark · The provider · Cyber
In Project Fortify, Google’s autonomous system spanning scanning, triage, and remediation, Carbon hit 100% recall on scanning and 95% precision on triage, on our internal benchmark.
Google DeepMind · 10 November 2026 · Read the source - SEC-bench Pro · The provider · Cyber
We evaluate on the May 2026 version of the benchmark: 183 historical JavaScript engine vulnerabilities in Chromium (V8) and Firefox (SpiderMonkey).
Google DeepMind · 10 November 2026 · Read the source - Scheming honeypot evaluations (Code Review + Agentic Coding) · The provider · Deception & sabotage
0.0% clear or borderline sabotage across all 3,021 runs and at most 0.2% scheming reasoning (1/511 borderline runs under default Agentic prompts, 0.0% Mild or clear
Google DeepMind · 10 November 2026 · Read the source - Sexually Explicit Material Safe Response Rate · The provider · Harmful content
Gemini 4 Carbon improves on the slight regression in Sexually Explicit Content Safety from Argon and saturates our single-turn evaluation.
Google DeepMind · 10 November 2026 · Read the source - Simulated cyberattack testing with ExploitBench and CyScenarioBench against an adaptive red-teaming agent · The provider · Misuse & jailbreak
In simulated cyberattack testing environments seeded with ExploitBench and CyScenarioBench, Carbon blocked 99% of attack trajectories from our adaptive red-teaming agent that can iteratively revisit its attack plans after each block.
Google DeepMind · 10 November 2026 · Read the source - Single-turn hate speech & harassment safety evaluation · The provider · Harmful content
Gemini 4 Carbon demonstrates improvement in Hate & Harassment Safety, both in single-turn and multi-turn evaluations, over Argon’s strong baseline.
Google DeepMind · 10 November 2026 · Read the source - Text to Text Safety · The provider · Harmful content
Reported result: +3.50pp versus Gemini 4 Argon
Automated content safety evaluation measuring adherence to safety policies across text prompts +3.5pp
Google DeepMind · 10 November 2026 · Read the source - Tone · The provider · Harmful content
Tone Automated evaluation measuring objective tone of model refusal +1.10pp
Google DeepMind · 10 November 2026 · Read the source - Tone evaluation · The provider · Harmful content
Gemini 4 Carbon shows similar performance in tone quality compared to existing models.
Google DeepMind · 10 November 2026 · Read the source - Unjustified-refusals · The provider · Misuse & jailbreak
Unjustified-refusals Automated evaluation measuring model’s ability to respond to borderline prompts while remaining safe +1.10pp
Google DeepMind · 10 November 2026 · Read the source - Unjustified-refusals · The provider · Misuse & jailbreak
Reported result: +0.10pp versus Gemini 4 Argon
Automated evaluation measuring over-refusals on benign or borderline prompts that can be safely fulfilled +0.10pp
Google DeepMind · 10 November 2026 · Read the source - agent monitoring and response platform alert comparison · The provider · Deception & sabotage
In RL training, high-severity monitor flags were 3.5x lower (0.007% vs 0.025% of rollouts) and inverse-probability-weighted confirmed alerts were 4.5x lower.
Google DeepMind · 10 November 2026 · Read the source - alignment-faking and training-gaming rubric audit of RL rollouts · The provider · Deception & sabotage
Neither judge flagged any rollout (0 / 100,000) for explicit training-update awareness – reasoning about how RL updates or data selection shape future weights (Level 3) – or strategic alignment faking or sandbagging to steer weight updates (Level 4), placing a 95% one-sided upper bound of 0.0043% on overt alignment-faking deliberation after re-weighting for stratified sampling (tightening Argon's 0.05% upper bound roughly tenfold; Figure 3.7).
Google DeepMind · 10 November 2026 · Read the source - cheating run audit · The provider · Deception & sabotage
Across alternative readings of ambiguous boundary cases, Carbon's cheating rate ranges from 0.0% to 4.1% (vs 0.5% to 9.5% for Argon).
Google DeepMind · 10 November 2026 · Read the source - condensed version of our alignment evaluations · The provider · Deception & sabotage
We ran a condensed version of our alignment evaluations to verify that model updates did not introduce new concerning misalignment behaviors.
Google DeepMind · 10 November 2026 · Read the source - internal-codebase honeypot runs · The provider · Deception & sabotage
Across 100,000 audited RL rollouts, low-friction instrumental-convergence scenarios, and 3,021 internal-codebase honeypot runs, we also observed zero overt alignment faking, self-preservation, or unprompted sabotage.
Google DeepMind · 10 November 2026 · Read the source - misalignment-monitor alert audit of RL training and internal coding deployments · The provider · Deception & sabotage
3.5x lower high-severity flag rate (0.007% vs 0.025% of rollouts) and 4.5x lower inverse-probability-w eighted confirmed alert rate in RL training; 7.2x lower confirmed alert rate per session (0.0031% vs 0.023%) and 10.2x per model call (0.073 vs 0.75 per 100,000 calls) in a matched internal coding cohort (~5x–6x lower as Carbon ramped more broadly); 0 model-initiated attacks on monitoring or logging infrastructure outside the task sandbox
Google DeepMind · 10 November 2026 · Read the source - neutrality evaluation · The provider · Bias
Gemini 4 Carbon shows similar performance to Argon on our neutrality evaluation.
Google DeepMind · 10 November 2026 · Read the source - normal-operation transcript checks · The provider · Deception & sabotage
Setting aside turns triggered by harness messages, safety-filter blocks, or user formatting requests leaves unprompted CoT-monitoring awareness at or below 0.002% on both models.
Google DeepMind · 10 November 2026 · Read the source - self-harm · The provider · Harmful content
& self-harm : evaluating whether model responses align with evidence-based suicide prevention best practices, refuse self-harm instructions, avoid validation of urges to self-harm, and surface relevant crisis resources.
Google DeepMind · 10 November 2026 · Read the source - static cyber jailbreak suite · The provider · Misuse & jailbreak
Gemini 4 Carbon achieves <1% QSR (0.45%) and 6% OSR across our expanded 50-objective static cyber jailbreak suite, indicating strong resistance to drop-in template jailbreaks.
Google DeepMind · 10 November 2026 · Read the source - CWE-bench · The provider · Cyber
Carbon reaches 71.0% on Collinear’s CWE-bench, turning validated findings directly into working code fixes.
Google DeepMind · 28 October 2026 · Read the source - ExploitGym · The provider · Cyber
Carbon sets a new state-of-the-art on the popular exploitation benchmark ExploitGym (54.8%).
Google DeepMind · 28 October 2026 · Read the source - Gray Swan’s Indirect Prompt Injection (IPI) Benchmark · The provider · Injection & agentic
While Argon previously led the Gray Swan’s Indirect Prompt Injection (IPI) Benchmark in prompt injection resistance with a very low 0.7% attack success rate, Carbon achieves 0.0% — resisting every evaluated attack attempt.
Google DeepMind · 28 October 2026 · Read the source - SEC-bench Pro · The provider · Cyber
Carbon achieves a state-of-the-art 88.2% on SEC-bench Pro, autonomously uncovering complex flaws buried in real-world software.
Google DeepMind · 28 October 2026 · Read the source
No document we read states: a release date. No incident on record names this version.
Some documents we hold are stored in lower case only, so their quotes are shown in lower case. The source link opens the original.
Previous version: Gemini 4 Argon · Compare the two
Documents listed for this version
- Gemini 4 Carbon — Safety & Security Report · Google DeepMind · 10 November 2026
- Gemini 4 Carbon — Model Card · Google DeepMind · 10 November 2026
- Gemini 4 Carbon · Google DeepMind · 10 November 2026