Brief IA
›

Gemini 4 Carbon: Google Advances Its Cybersecurity Results

Research·Tom Levy·

Gemini 4 Carbon: Google Advances Its Cybersecurity Results

Gemini 4 Carbon: Google Advances Its Cybersecurity Results
Key Takeaways
1Google announces a 100% recall in an internal evaluation of Fortify with Carbon.
2The API costs $2 per million incoming tokens and $10 per million outgoing tokens at launch.
3Axiom Bio is transferring its scientific analyses and agents to Carbon and Argon.
Why it matters : /swarm offers to distribute a task among multiple agents, using Carbon for deep reasoning tasks.

Google announces a 100% recall for Carbon during an internal evaluation of its Fortify system. The model is now accessible via API to paying customers, while its rollout to Google AI Ultra subscribers continues. Its announced applications range from scientific research to coordinated agent teams in Antigravity.

Google claims vulnerabilities discovered at SaaS providers

In Scan for Good, a program by Wiz focused on public digital infrastructures, Google reports that Carbon identified 0-day vulnerabilities at two major SaaS providers in less than two hours. The company states that this operation reduced what would have taken several weeks of work to just one afternoon. Wiz collaborated with the affected organizations to fix the vulnerabilities in record time, thereby limiting the potential impact on tens of thousands of businesses.

The results from Fortify, Google's autonomous system covering scanning, sorting, and remediation, come from an internal evaluation: Google announces a 100% recall for scanning and 95% precision for sorting. On open-source repositories, the company claims a relative improvement of over 25% in recall compared to Argon, with the detection of logical defects that were previously unidentified. Google states that it is one of the leading users of open-source software and believes these advancements benefit both its own systems and the entire ecosystem.

Another example: Google's CVR team, which specializes in cloud vulnerabilities, used Carbon to demonstrate the exploitability of two bugs present in widely used open-source software. Google assesses the severity of these vulnerabilities to be comparable to that of the PHP2Shell exploit and indicates that a proof of concept created with Carbon led maintainers to reclassify the bug as critical and publish a fix without delay. Technical details will be shared after the responsible disclosure process is completed.

Google describes encouraging practical feedback, both internal and external, for this version, which it presents as the most advanced in cybersecurity. Through Fairwind, it claims strong performance on SEC-bench Pro for detection, ExploitGym for exploitation, and CWE-bench v1 for vulnerability remediation. Software inspection, analysis of defect chains between systems, and generation of source-verified patches are among the announced capabilities.

Axiom Bio transfers its analyses to Carbon and Argon

Axiom Bio, a company specializing in drug toxicity, considers Carbon to be the most effective model in translational sciences and credits it with discovering toxicity mechanisms that it had not previously identified. The transfer of its scientific production analyses and agents to Carbon and Argon is underway, with the goal of designing safer drugs.

Google claims an improvement in Carbon's reasoning in physics, mathematics, and life sciences, to interpret experimental data, test hypotheses, and solve multi-step problems. Its researchers have used Carbon in Antigravity for enzymatic assays, with only lab notebooks and raw data as inputs. According to Google, the model highlighted difficult-to-interpret anomalies and diagnosed experimental artifacts, including an incomplete automated mix in deep wells, which improved reproducibility. Google also mentions better methods for deconvoluting chromatographic peaks and detecting trace samples.

For software verification, researchers designed AlphaProof Code and applied it to the cryptographic protocol code of Signal. Carbon, combined with other models, was used to translate security objectives into mathematical theorems; agents then generated millions of lines in Lean, a language for mathematical proofs. Google reports that key security properties were proven in a matter of weeks, whereas this work historically took years for specialized teams. The company sees potential for scaling formal verification without generalizing this result to all software.

Google charges Carbon at Argon's launch rate

Access to Carbon via API is offered through Google AI Studio and Gemini for businesses, available to all paying developers and companies. The launch rate is the same as Argon's: $2 per million incoming tokens and $10 per million outgoing tokens. Deployment to all Google AI Ultra subscribers is ongoing in Antigravity, the Gemini app, and the AI mode of search.

Le brief IA que lisent les pros

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

Carbon is accessible to developers, businesses, scientists, and consumers. It succeeds Gemini 4 Argon and was initially offered, two weeks before this announcement, to a group of cybersecurity advocates through the Fairwind program. According to Google, Carbon represents its most advanced model for long and complex agentic tasks, while Argon remains preferred for the majority of uses in software engineering and knowledge management. Use cases for Carbon include sophisticated code development, debugging, or open scientific research, and it utilizes a larger volume of tokens than Argon to allow for deeper exploration, reasoning, and verification.

In Antigravity, Google highlights the /swarm command: a single instruction allows for the launch of a coordinated team of agents, whose activities are tracked live. Carbon can be assigned to architecture, audit, or coordination missions, alongside faster and lighter models operating in isolated spaces in parallel. Adversarial verifiers test each modification.

Google cites among the targeted uses multi-package migrations, root cause investigations between subsystems, and complex architectural overhauls. For web applications, the company describes a functioning that alternates between construction, error management, and checks of both code and the displayed interface.

Carbon ranks second in the MKA multimodal evaluation

Google and Artificial Analysis launched the Multimodal Knowledge Agent (MKA) benchmark to evaluate multimodal capabilities in agentic tasks. Google believes that existing benchmarks do not accurately measure agentic reasoning on dense files combining images, graphics, tables, plans, and diagrams. In MKA, Carbon ranks second, with an average cost per task over 6X lower than that of the model that came in first.

On CursorBench 4.0, WANDR from Perplexity, and AutomationBench from Zapier, Google ranks Carbon in the top position. These evaluations are described as independent and representative of concrete tasks, with a cost per task more than twice lower than that of the next best external model. According to Google, initial financial analyses indicate performance equivalent to that of the best model tested at Hebbia on E2E Financial Agentic.

Carbon also ranks first among 23 configurations on General Financial Intelligence from Balyasny Asset Management, according to Google. This test covers law, the use of agentic tools and their execution, programming, data engineering, as well as contextual perception and understanding. For legal, financial, and business operations, Google describes a model combining document search, visual understanding, and reasoning over large datasets. It also attributes the production of tables, presentations, and reports subject to strict formats to it.

/swarm demonstrations range from building to 3D gaming

Architectural demonstrations include Google Platform 37: seven public PDFs of municipal planning, from which 206 technical views were extracted, enabled interactive reconstruction. Google announces 906K triangles in Three.js, following a prompt and 10 overnight loops with /swarm. Another team of agents created, also from a single prompt, an interactive 3D exhibition on the construction of the Eiffel Tower between 1887 and 1889, utilizing historical photographs and engineering drawings.

In the scientific domain, Google presents a multiscale textile visualizer obtained in Antigravity from academic work, with minimal instructions. It models virtual fibers using finite elements, with 3D wire meshes, polar tracking of 110 fibers, and dynamic stress-strain curves. Another technical achievement: a Stein-Schule fortepiano from 1780, built from CAD plans and X-rays, playable on 61 keys with escape physics. The instrument is integrated into a 3D representation of Koerner Hall at the Royal Conservatory of Music in Toronto, with spatial acoustics dependent on seating.

Google also attributes visual and spatial advancements applied to gaming and creative coding to Carbon. An open-world game features a pelican on a GBike in the Mountain View campus: its sub-agents used Blender with Omni and Nano Banana references, Three.js, and WebGL shaders, then Web Audio for procedural sound generation and Lyria for musical composition. An underwater sequence of Atlantis begins with a video produced with Gemini Omni; its procedurally generated ruins, volumetric caustic lighting, and water simulation in Blender are then exported to Three.js for real-time rendering in the browser.

On the commercial side, Google reports a redesign of the Google Merchandise Store achieved in two prompts, with five artistic directions and product visuals generated by Nano Banana 2.1. This achievement mobilizes front-end design and motion design skills in Antigravity. Online versions of the store, the game, and the fortepiano are offered, along with a front-end design skill in Gemini.

Brief IA — L'actualité IA en français

L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.