Gemini 4 Carbon tops a model index—but remains limited to selected cyber defenders
Google reports strong vulnerability-finding, exploitation and patching scores. One independent composite ranks Carbon first, while public access, reproducible safety evidence and operational outcomes remain limited.
The short version
- What happened
- Artificial Analysis gives Gemini 4 Carbon a score of 59, the highest published result on its nine-evaluation composite, one point above the next listed model.
- Who it affects
- Security teams could gain faster help finding and repairing vulnerabilities, while more capable exploitation also raises misuse concerns. Controlled access means most defenders cannot yet test whether the model improves their own detection quality, patch safety or response time.
- What remains unproven
- No public API
Keep in mind
No public API
Selected cohort only
Cyber scores reported by Google
Evaluations without guardrails
Internal safety tests lack denominators
No Carbon-specific detailed safety report
No incident-response outcome study
Source analysis · 2 linked sources. Primary analysed source date: 28 October 2026.
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Read the full report · 3 min The evidence, methods, implications and limitations
The strongest public signal is a narrow benchmark lead
Google's new cyber-focused frontier model leads the current Artificial Analysis Intelligence Index with 59. The independent composite averages agents, coding, general capability and scientific reasoning at 25% each. A one-point lead is useful evidence of broad benchmark strength, not proof of a decisive advantage in operational security work.
Google separately reports Carbon at 88.2% for vulnerability finding on SEC-bench Pro, 54.8% for exploitation on ExploitGym and 71.0% for patching on CWE-bench. It says the evaluations ran without guardrails, a condition that may differ from deployed use.[1][2]
Controlled access leaves the safety case unfinished
Google reports lower cheating, unsupported claims, overreach and prompt-injection success than earlier models, but publishes the Carbon figures in a launch post rather than a new detailed technical or safety report. Internal alerts and live-coding comparisons lack denominators and external replication.
Starting this week, Carbon is going to selected cyber defenders before a gradual trusted-partner rollout. That caution may reduce misuse exposure, but it also limits independent testing. Evidence that would change the assessment includes a full model card, reproducible guarded evaluations, transparent access criteria and measured defensive outcomes in real incidents.[1][2]
The distinction that matters
Three different kinds of evidence
The launch combines an independently run composite, company-reported cyber benchmarks and internal safety observations.
Independent
Nine-evaluation composite
Score 59; current published leader
Company reported
Three cyber benchmarks
Strong scores without public guarded replication
Operational
Selected-defender rollout
Real-world outcomes not yet reported
Watch the background, then check the evidence
Illustrative photo · not a video frame
The Royal Society · 2 min 31 sec · Background explainer
What is artificial intelligence? | The Royal Society
YouTube loads only when you choose; playback starts when you press play.
A short introduction to artificial intelligence from the Royal Society. Start here for basic concepts, then use current reporting to explore what has changed.
Evidence check: Published in 2018, before today’s generative AI tools. Useful for basic concepts, rather than an overview of current capabilities.
Published 2018-11-28. Source checked 14 October 2026. Check captions on the publisher’s video page.
Explore current AI technologyPhoto: ThisisEngineering / Unsplash · Licence
Continue the story
Read next
Continue exploring cybersecurity: compare the finding and its evidence limits.
Google packages faster reasoning with a cyber-focused Gemini variant
Google says Gemini 3.8 Flash improves reasoning and coding at the speed and price point of its previous workhorse model, alongside a specialised Flash Cyber version for defensive security tasks.
Read the next reportExplore the full source and evidence record
What this means for people
- Security teams could gain faster help finding and repairing vulnerabilities, while more capable exploitation also raises misuse concerns. Controlled access means most defenders cannot yet test whether the model improves their own detection quality, patch safety or response time.
Global context
The initial rollout uses Google's Fairwind trusted-partner programme and U.S. voluntary pre-release testing. Access, legal oversight and cyber-defence capacity differ widely across countries.
What the evidence does not yet show
- No public API
- Selected cohort only
- Cyber scores reported by Google
- Evaluations without guardrails
- Internal safety tests lack denominators
- No Carbon-specific detailed safety report
- No incident-response outcome study
What to watch next
- Public API or broader access
- Carbon model card
- Independent guarded cyber evaluations
- Misuse testing
- Patch regression rates
- Real-world defender outcomes
Living evidence record
Impact record IAI-1G8UM36
Evidence stage
Announced
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
28 October 2026
Source trail
2 direct sources across 2 source types.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Related-source reporting disclosure
This record analyses 2 linked source records around the same underlying development. The extra records add method, date or context, but they do not by themselves constitute independent replication of every performance claim or predicted outcome.
Evidence trail
Sources used for this report
Links checked 28 October 2026
Research question, topics and regions
Topics: Cybersecurity · Frontier models · Vulnerability research · Model safety · Benchmarks · Controlled rollout
Regions: United States · Global
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.