EXPLORINEER
en

JP — ITmedia

Google Unveils "Gemini 4 Carbon," Strengthening Cyber Defense Two Weeks After "Argon" Launch

AI1 min read

Google has announced Gemini 4 Carbon, a new frontier AI model that it says is its most capable so far. It is strongest at cybersecurity and long, complex tasks, and it is first going to a small group of cyber defense organizations.

What happened

On October 28 (US time), Google unveiled Gemini 4 Carbon. Google says it is the most powerful AI model it has built. It is better at finding and fixing computer security weaknesses and at handling long, complicated jobs. For now, only some cyber defense groups can use it. Google plans to open it to developers, companies, and general users later, step by step.

Key numbers and facts

  • Score on the Artificial Analysis Intelligence Index (v5.0), a test of AI ability: Carbon 59 points, first place. Argon 50 points. That is 9 points higher.
  • Google's internal tests: safe code generation success rate more than doubled compared with Argon.
  • Hallucinations (made-up claims) fell to about half of Argon's level.
  • Serious alignment problems found in coding settings fell by about 80%.
  • Success rate of indirect prompt injection attacks (Gray Swan IPI): 0.7% for Argon, 0.0% for Carbon.

Background

Google released Gemini 4 Argon on September 30 and made it generally available on October 13, only about two weeks before Carbon. Logan Kilpatrick, a product leader at Google DeepMind, said Carbon is a big step up from Argon. Google also worked with CAISI, an AI testing group under the US National Institute of Standards and Technology (NIST), on safety checks before release.

Why it matters

Fast model releases show how quickly AI is improving. Strong cyber skills can help defenders find weak spots in software. The same skills could also be misused by attackers. That is why Google is releasing Carbon slowly and checking safety at each step.

What to watch

  • When Carbon becomes available to developers, companies, and general users. Google has not given a date.
  • Results from the Fairwind Program, where cyber defense groups test the model.
  • Whether independent safety checks confirm Google's results on hallucinations and prompt injection.

Glossary

Gemini 4 Carbon
Google's new AI model, made for strong cybersecurity and long, complex tasks.
Google DeepMind
Google's AI research group that builds models like Gemini.
Artificial Analysis Intelligence Index
A test that scores how well AI models perform across many tasks.
CAISI
An AI testing group under the US National Institute of Standards and Technology (NIST) that checks advanced AI models for risks.

Same story, different eyes

Ad

Add 150 words. Learn 10 a day.

Dicter word books send only the words you learn to your review notes.

More in this topic

AIheise online

New LLM: Gemini 4 Carbon Is an Exploit Expert

Google has unveiled Gemini 4 Carbon, its newest large language model, which succeeds Argon, announced four weeks earlier. Like Argon, Carbon is initially restricted to US intelligence agencies and selected users, with wider access planned gradually through Google's Fairwind program; Google says it has not yet been able to carry out a successful prompt injection against the model. The biggest gain over Argon is in exploit development, with Carbon scoring nearly 55 percent on the Exploitgym benchmark, up from 42 percent.

AIVnExpress

Disillusioned by Rushed AI Adoption

Many companies worldwide are struggling to clean up the mess caused by rushing to adopt AI tools that are not yet stable, leading to lost employee time, information leaks and higher costs. Reuters reported that Meta paused plans to cut staff and expand AI use because employees spent too much time fixing technical and security errors. Business Insider reported that Walmart store and warehouse workers faced unreasonable demands, system errors and false alerts from AI tools used for task allocation, scheduling and performance tracking.

AITechCrunch

Anthropic just launched tools to show companies what their AI spending is buying

Anthropic has launched new reporting tools that let enterprise managers see a team's output over a chosen date range, what it cost, and where workflows broke down, using internal analytics and employee surveys. The company says it won't use this data for model training by default, and the reports will include recommendations for improvement. Anthropic is also creating pooled group budgets to help teams manage token spending amid growing concern among large customers about return on investment.

AIThe Register

Anthropic engineer declares the coding interview dead

Anthropic engineering manager Felix Rieseberg says he has stopped using traditional coding interviews for his team, arguing that Claude now outperforms candidates on such tests. Jaana Dogan, a principal engineer at Google, expressed a similar view, while labor economist Nicole Bachaud said the broader tech industry is not yet abandoning skills assessments.