Gemini 4 Argon AI Enhances Security: Resolves 12 Simulated Attacks, Compared to 5

Google has launched Gemini 4 Argon AI, a model with stronger cybersecurity capabilities and real-time protections.

During limited access testing, Wiz used it to find a flaw in hospital software that previous models had not detected. This defect concerned the protection of hundreds of thousands of medical records. The key now is to understand whether the defenses can keep pace with the new capabilities.

Key Points

• Argon has completed 12 simulated cyberattacks, compared to five from the previous generation.
• Three levels of protection filter requests, check internal signals, and monitor responses.
• In external biological jailbreak tests, less than 1% of subtasks were unlocked.
• In the Gray Swan benchmark on prompt injection, attacks succeeded in 0.7% of cases.

The Security of Gemini 4 Argon AI Reaches Real Systems

The model has already been employed on real software, not just in the lab. According to Google DeepMind, which explained it in a blog post, it improves complex work in programming, finance, law, cybersecurity, and science.

Advancements in coding and reasoning have helped security teams identify and fix vulnerabilities. Besides the Wiz case, the company's teams have patched bugs in internal and open-source software. Among these are fundamental vulnerabilities affecting systems worldwide and are undergoing responsible disclosure.

The launch took place on October 13, 2026, reported The Verge. The outlet also cites internal uses beyond security: the model has halved the computing power required by the YouTube comments section. It also helped correct errors in the physics simulator for drones used by Wing, Alphabet's drone delivery company.

In biology assessments, unprotected versions showed greater capabilities in guiding, step by step, the resolution of problems in complex laboratory procedures. For this reason, the company has applied protections to both cybersecurity and biology, as part of its Frontier Safety Framework.

From the Fairwind Program to Broader Access

Before the general rollout, Argon transitioned from a phase reserved for a few users. Cyber defenders tested it through the Fairwind Program, while internal security teams analyzed critical systems inside and outside the company.

Now Google is expanding access and bringing the Fairwind program to more trusted defenders. According to launch coverage, the model is available to Pro and Ultra subscribers via the Gemini app and the AI mode of Search. Ultra subscribers also find it in Antigravity, the company's agent development platform.

At launch, the price is $2 per million input tokens and $10 per million output tokens. The rates double at the end of the introductory period, at the end of November.

The same coverage notes that a lower price per token does not necessarily mean a lower cost for tasks. Different models consume a different number of tokens to perform the same work.

Three Levels Against Abuse and Jailbreak

Argon's protection system combines three tools: request filtering, internal model analysis, and continuous monitoring. The goal is advanced AI protection against those who mask harmful requests or fragment them into long conversations.

In the first level, smaller models examine incoming requests and the context of the conversation. In the second, some activation probes analyze internal neural signals, looking for knowledge related to cyberattacks or biological threats.

The third level is continuous monitoring, which combines those signals as responses take shape. It can stop an agent mid-response.

Google reported that violations with harmful requests have nearly halved on an adversarial test set. In external biological jailbreak tests, less than 1% of evaluated biological subtasks were unlocked. Protections, faster jailbreak fixes, and monitoring of repeated abuses contributed to this.

Cybersecurity Results and Agent Monitoring

In the company's end-to-end tests, Argon resolved all 12 scenarios of simulated cyberattacks. The previous generation completed five.

On the Indirect Prompt Injection benchmark from Gray Swan, resistance to prompt injection led to attack success in 0.7% of 15 adaptive attempts. The result comes with adversarial training and real-time security checks.

Real-time security monitoring also involves an automatic system against misalignment. It reads agents' actions and their internal reasoning, blocking serious violations of limits and reporting serious cases to human responders. It operates in products, internal programming agents, training sessions, and capability assessments.

Training focuses on three behaviors: respecting system boundaries, honestly reporting negative results, and asking for confirmation from a person. Attempts at shortcuts or cheating have dropped to 2% in coding benchmarks. Honeypot tests found no hidden behaviors, such as sabotaging security monitoring.

The company avoids penalizing internal thoughts during training. In an accidental experiment, these penalties reduced openly discussed shortcuts by 61%. However, those taken silently rose from 11.2% to 36.4%.