US Europe Arab Asia Africa | Politics Economy Oil&Energy Entertainment Sport French Spanish

Google Reserves Gemini 4 Argon For Cybersecurity Partners Arabian Post

2026-10-01 02:21:19

(MENAFN- The Arabian Post) clearfix">Google has announced Gemini 4 Argon, its most capable artificial intelligence model, with initial access restricted to cybersecurity partners as it challenges OpenAI and Anthropic in software engineering and complex professional work.

The September 30 announcement introduces a new flagship for Google's Gemini range, but leaves consumers and most developers waiting. The company has not specified a date for wider availability, saying it will gather feedback and strengthen safeguards before expanding access.

Trusted cybersecurity specialists are receiving Argon through Google's Fairwind programme. Google said it was also participating in the US government's voluntary process for access to models before their public release, placing security evaluation at the centre of the rollout.

<p>Koray Kavukcuoglu, Google's chief AI architect and a senior vice-president at Google DeepMind, said the model was designed to sustain reasoning through lengthy tasks involving software development, legal and financial work, and cyber defence.</p><p>Google plans to begin its broader release with paying customers using its application programming interface and subscribers to Google AI Ultra. Those plans distinguish the announcement from an immediately accessible launch across the Gemini consumer service. blog. google</p><p>A company spokesperson described Argon as larger than Google's previous advanced Pro models and comparable with OpenAI's Astra and Anthropic's Opus on important coding and cybersecurity tests. Google also confirmed it no longer planned to release Gemini 3.5 Pro, which chief executive Sundar Pichai had earlier indicated would arrive in June.</p><p>The performance picture is mixed. Google's published comparisons put Argon ahead on several tests, but behind competitors on others, including two of the four coding benchmarks in its announcement. The figures are company assessments rather than evidence that Argon leads every category of professional AI use. reuters. com</p>See also  Apple raises evidence-destruction claims against OpenAI <p>On DeepSWE v1.1, a test of extended software engineering tasks, Google gave Argon a score of 77.9 per cent, compared with 74.2 per cent for Claude Opus 5.5 and 74.1 per cent for GPT-6 Astra.</p><p>Google also claimed leading results in financial research and legal drafting. It reported 51.3 per cent on Zapier's AutomationBench, which assesses execution across business functions, and 91.7 per cent on LVBench, a test of understanding long videos.</p><p>Argon's maximum output is being increased to one million tokens from the previous 64,000. Tokens are the units of information processed and generated by language models. Google said the expanded allowance supports longer reasoning sequences and more demanding tasks within a single run. 9to5google. com</p><p>Cybersecurity is the immediate deployment priority. Google said Argon could find, validate and repair critical software vulnerabilities autonomously. Trusted defenders and Google's internal teams will receive access without cyber guardrails, allowing them to use the model's full defensive capabilities.</p><p>Google said security company Wiz was already using Argon through Scan for Good, an initiative offering protection for critical public infrastructure. It said the model identified a serious weakness exposing personal information in healthcare software used by hospitals worldwide.</p><p>The announcement did not identify the affected software or provide technical particulars of the vulnerability. Google's claim that earlier frontier models missed the weakness therefore remains a company account of the demonstration. blog. google</p><p>Before wider distribution, Google said it was reinforcing protections against malicious cyber activity and chemical, biological, radiological and nuclear misuse. It is also addressing prompt injection, where hostile instructions embedded in material an AI reads attempt to redirect its behaviour.</p>See also  Snapchat adds Plans tool for organising meet-ups <p>Other controls monitor the model's reasoning and actions and can halt execution if it goes beyond a user's intentions. Google said internal and external specialists had tested safeguards through simulated attacks, while its testing environments were being further isolated. blog. google</p><p>Inside Google, Argon is supporting debugging and extensive code migrations. The company said projects included moving more than 800,000 lines associated with the Fuchsia operating system's Zircon kernel towards Rust, with automated and manual audits before production deployment.</p><p>For its libgav1 video decoder, Google said Argon replaced 32,000 lines of specialised code in an existing Rust version, producing identical video output while running 2.7 times faster than that earlier version in testing.</p><div class='code-block code-block-1'><h2></h2></div>

MENAFN01102026000152002308ID1111747544

Legal Disclaimer:
MENAFN provides the information “as is” without warranty of any kind. We do not accept any responsibility or liability for the accuracy, content, images, videos, licenses, completeness, legality, or reliability of the information contained in this article. If you have any complaints or copyright issues related to this article, kindly contact the provider above.

Google Reserves Gemini 4 Argon For Cybersecurity Partners Arabian Post

More Story