Google launches Gemini 4 Carbon, an AI model claiming top scores in vulnerability discovery, exploitation and patching
Matéria consolidada. Reúne 2 matérias que o Tifon tinha publicado separadamente sobre a mesma história. A IA reescreveu o relatório com o que as 2 fontes trouxeram em 30/10/2026 00:06; os links antigos trazem para cá.
Google unveiled Gemini 4 Carbon on 2026-10-29, a frontier model that it says leads rivals at finding, exploiting and patching software vulnerabilities. Google reports 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench (thehackernews.com, securityweek.com). Access is initially limited to selected cyber defenders and will expand through the Fairwind Program, which reportedly includes over 650 organizations (securityweek.com). Google also submitted the model to the U.S. government's voluntary pre-release testing process (thehackernews.com). Carbon follows Gemini 4 Argon, released two weeks earlier (securityweek.com). The launch came alongside OpenAI's release of GPT-6.1 Terra, which meets the Critical cybersecurity threshold under its Preparedness Framework, and an Anthropic disclosure that an unreleased internal research model found critical flaws in widely used software (thehackernews.com). The benchmark figures are Google's own and have not been independently verified. Together these events show AI-assisted exploit development moving toward near-autonomous capability, which creates both defensive opportunity and misuse risk.
Entenda a falha
A vulnerabilidade
The only CVE named is CVE-2026-33096, a denial-of-service flaw in the Windows HTTP.sys driver, CVSS 7.5. The articles give no attacker prerequisites, affected versions or patch identifiers, only that Microsoft has fixed it. The same Anthropic disclosure lists a one-click remote code execution in Google Chrome, a one-click exploit chain in Apple Safari that could expose browsing history, cookies and other sensitive data, and an Android lock screen bypass. None of these has a CVE ID in the articles, and Google, Apple and Microsoft are reported to have fixed all of them.
Como funciona
The articles do not describe how any of these flaws work internally. They do not say what input reaches HTTP.sys, the browsers or the Android lock screen, or at which step processing fails. The only details are the user-facing effects: the Chrome and Safari issues need one click from the victim, and the HTTP.sys flaw ends in a denial of service.
Por que é vulnerável
The articles do not state the root cause of any of the flaws, why they went unnoticed, or what each fix changes. They report only that Google, Apple and Microsoft fixed the issues.
Onde está
The only component named is the Windows HTTP.sys driver for CVE-2026-33096. No file, function, commit or public advisory is given.
Como foi descoberta
According to The Hacker News, Anthropic disclosed earlier this week that an unreleased internal research model uncovered these flaws in widely used software. Anthropic says such models can now build multi-step exploits in open-source and proprietary software with little human involvement, and prove them with a working exploit that anyone can run and verify. The articles do not say when the flaws were reported to the vendors or how the disclosure was coordinated, only that Google, Apple and Microsoft have fixed them.
Explicação gerada por IA a partir da matéria e dos artigos, em 30/10/2026 00:21.
Detalhes técnicos
- Cyber benchmarks: In Google's own comparison, with no guardrails applied, Carbon beat OpenAI GPT-6 Astra, Anthropic Claude Opus 5.5 and Claude Fable 5.1 (securityweek.com). SEC-bench Pro measures real-world flaw discovery, and CWE-bench is a Collinear patching benchmark (securityweek.com). Wiz observed Carbon leading web penetration testing benchmarks (securityweek.com).
- General index: Carbon scores 59 on the Artificial Analysis Intelligence Index, ahead of Claude Opus 5.5 (58), GPT-6 Astra (57) and GPT-6.1 Sol (55) (thehackernews.com).
- Secure code and patch quality: Google says Carbon writes secure code twice as often as Argon (both sources). Google's Chrome team found its patch quality more than 10x better than a rival frontier model (securityweek.com).
- Prompt injection and misalignment: In Gray Swan's indirect prompt injection benchmark no attack succeeded, versus 0.7% for Argon. On Terminal-Bench 4.0, cheating attempts fell to 0.4% and none succeeded (securityweek.com). Google also says Carbon seeks human input under ambiguity (thehackernews.com) and has lower misalignment rates (securityweek.com). OpenAI reports similar prompt-injection resistance for GPT-6.1 Terra (thehackernews.com).
- Pricing: $2/$10 per million input/output tokens, rising to $4/$20 after the introductory period (securityweek.com).
- Anthropic's internal model: It reportedly produced a one-click remote code execution in Google Chrome, a one-click exploit chain in Apple Safari exposing browsing history and cookies, an Android lock screen bypass, and a denial-of-service flaw in the Windows HTTP.sys driver (CVE-2026-33096, CVSS 7.5). Google, Apple and Microsoft have fixed all of these (thehackernews.com).
- Google's framing: Google positions exploitation capability as a defensive aid for prioritizing remediation (securityweek.com).
Remediação
- Apply vendor patches for CVE-2026-33096 and the referenced Chrome, Safari and Android fixes (thehackernews.com).
- Expect faster vulnerability discovery by adversaries using frontier models. Prioritize patch velocity and exploit-aware detection (thehackernews.com).
Relatos divergentes
- Competitor comparison: thehackernews.com lists Claude Opus 5.5, GPT-6 Astra and GPT-6.1 Sol in the Intelligence Index comparison. securityweek.com lists GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1 in Google's cyber benchmark comparison. These may be different comparisons rather than a true contradiction, and both rely on Google's claims.
- Fairwind scope: thehackernews.com describes access as a subset of cyber defenders. securityweek.com says Fairwind reportedly includes over 650 organizations, including governments, critical infrastructure operators and large tech platforms. That figure is reported but unconfirmed.
- Prompt injection: thehackernews.com says Carbon resisted every evaluated attempt. securityweek.com gives the more detailed figure of zero successful attacks on Gray Swan's benchmark versus 0.7% for Argon.
Linha do tempo
-
Google unveils Gemini 4 Carbon, claiming top benchmarks in vulnerability discovery, exploitation and patching
ler a análise desta fonte ocultar
Google announced Gemini 4 Carbon, a frontier AI model it claims outperforms rivals at detecting, exploiting and patching software vulnerabilities. Access is initially limited to selected cyber defenders and will expand through Google's Fairwind Program, which gives governments, critical infrastructure operators and large tech platforms early access to advanced models and reportedly includes over 650 organizations. Carbon follows Gemini 4 Argon, released two weeks earlier. Google positions exploitation capability as a defensive aid for prioritizing remediation, and reports improved resistance to indirect prompt injection and lower misalignment rates.
Technical Details
Google reports Carbon scored 88.2% on SEC-bench Pro (real-world flaw discovery), 54.8% on ExploitGym and 71.0% on Collinear's CWE-bench patching benchmark, ahead of OpenAI GPT-6 Astra, Anthropic Claude Opus 5.5 and Claude Fable 5.1 in Google's own comparison, with no guardrails applied. Internal benchmarks claim Carbon is twice as successful as Argon at writing secure code, and Google's Chrome team found patch quality more than 10x better than a rival frontier model. Wiz observed Carbon leading web penetration testing benchmarks. In Gray Swan's indirect prompt injection benchmark no attack succeeded (vs 0.7% for Argon), and on Terminal-Bench 4.0 cheating attempts fell to 0.4% with none succeeding. Pricing starts at $2/$10 per million input/output tokens, rising to $4/$20 after the introductory period.
-
Google launches Gemini 4 Carbon with state-of-the-art offensive and defensive cyber benchmarks
ler a análise desta fonte ocultar
Google unveiled Gemini 4 Carbon, a frontier model claiming leading results on cybersecurity benchmarks: SEC-bench Pro (88.2%) for vulnerability discovery, ExploitGym (54.8%) for exploitation, and CWE-bench (71.0%) for patching. It scores 59 on the Artificial Analysis Intelligence Index, ahead of Claude Opus 5.5 (58), GPT-6 Astra (57) and GPT-6.1 Sol (55), and is said to generate secure code twice as often as Gemini 4 Argon. Access is initially restricted to a subset of cyber defenders under Google's Fairwind Program, with the model also submitted to the U.S. government's voluntary pre-release testing process. In parallel, OpenAI released GPT-6.1 Terra, which meets the Critical cybersecurity threshold under its Preparedness Framework, and Anthropic disclosed that an unreleased internal research model found critical flaws in widely used software. The trend confirms that AI-assisted exploit development is moving from specialist teams to near-autonomous model capability, raising both defensive opportunity and misuse risk.
Technical Details
Anthropic's internal model reportedly produced a one-click remote code execution in Google Chrome, a one-click exploit chain in Apple Safari exposing browsing history and cookies, an Android lock screen bypass, and a denial-of-service flaw in the Windows HTTP.sys driver (CVE-2026-33096, CVSS 7.5). All issues were fixed by Google, Apple and Microsoft. Google states Carbon resisted every evaluated indirect prompt injection attempt and seeks human input under ambiguity; OpenAI reports similar prompt-injection resistance for GPT-6.1 Terra.
Remediation
Apply vendor patches for CVE-2026-33096 and the referenced Chrome, Safari and Android fixes. Security teams should anticipate faster vulnerability discovery cycles by adversaries using frontier models and prioritize patch velocity and exploit-aware detection.