Google has unveiled Gemini 4 Argon and begun giving selected cybersecurity defenders access to its full cyber capabilities, while holding back a wider release for further safety testing. The September 30 launch puts an AI model that Google says can find and patch serious software flaws into approved hands first, with no date for general access.
The most striking example concerns hospitals. In its announcement, Google says Wiz used Argon to discover a critical vulnerability exposing sensitive personal information in healthcare software used by hospitals worldwide. Earlier frontier models had missed it, Google says. That is a concrete claim about finding a dangerous flaw, although the announcement does not establish that it has been fixed.
Access is through Google’s Fairwind Program, which works with more than 650 partners globally. That figure describes the whole network, not the number receiving Argon. Google has not disclosed how many organizations have the new model.
| Who | Access and conditions | What remains unspecified |
|---|---|---|
| Selected Fairwind defenders | Rollout underway, with organizational vetting, authenticated employee access and permitted defensive or research uses. | Cohort size and exact selection criteria. |
| Google’s internal teams | Google says they receive full cyber capabilities alongside trusted defenders. | Which model restrictions are removed. |
| Developers, enterprises and consumers | Later access promised, starting with paid API customers and Google AI Ultra subscribers. | Release date and safety thresholds for expansion. |
What the hospital finding establishes
Google attributes the discovery to Wiz’s Scan for Good initiative, which seeks out serious exposures in critical public infrastructure. But Google does not identify the affected software, vendor or type of vulnerability. It also supplies no severity calculation, names of the earlier models, disclosure date, patch status or account of whether any data was accessed. The published evidence establishes Google’s claim, rather than an independently auditable vulnerability finding.
Wiz’s own September 24 account of Scan for Good explains that model-generated hypotheses need expert validation and promises anonymized examples after remediation. It credits its tooling especially to Gemini 3.8 Flash Cyber and does not mention Argon or identify this hospital-software case.
That article predates Argon’s announcement, so it does not contradict later or unpublished work. Coordinated disclosure can also legitimately delay technical details while a vendor repairs a flaw. For now, the hospital example is a promising reported discovery, not evidence that hospitals have received a fix or that Argon consistently finds what competing models miss.
Full cyber capabilities, restricted users
The same ability to locate a security weakness can help someone repair it or exploit it. Google says it is strengthening defenses against cyber misuse before broad availability, alongside protections against malicious instructions hidden in material the model reads. It also describes monitoring the model’s reasoning and actions, with the ability to stop execution, and hardening the isolated computing environments used for testing.
Google says approved defenders and its internal teams receive Argon “without cyber guardrails.” Access still carries organizational, account and permitted-use restrictions. Google does not specify which refusal behaviors, monitoring systems or tool restrictions that phrase removes, so it should not be read as a claim that every safeguard disappears.
Fairwind requires phishing-resistant multifactor authentication and limits access to internal security, incident-response or penetration-testing teams. Organizations must track employee use, keep work within permitted defensive or research tasks, and prohibit sharing or resale. Those controls make who receives the model part of the safety strategy. Their effectiveness remains unmeasured publicly: the rules disclose no audit frequency or enforcement results.
As of October 3, Google’s model-card index had no Argon entry. The launch describes safety work but publishes neither Argon-specific red-team metrics nor the thresholds required for wider release. That limits outside scrutiny of the decision; it does not establish that evaluations or safeguards are absent.
Strong test results do not settle everyday performance
Google reports 77.9% on DeepSWE v1.1, a test of extended software-engineering tasks, and a joint-leading 68% on CWE-bench v1, which tests vulnerability repair. These are company-reported results on defined tasks, not success rates for arbitrary codebases or hospital systems.
Independent testing by Vals AI supports a strong but uneven picture. At high reasoning effort, Argon led the aggregate Vals Index with 68.90%, but ranked second on CyberBench, fifth on Terminal-Bench 4.0 and seventh of eight on CUA-bench, a computer-use evaluation. These suites test different skills; none verifies Wiz’s hospital finding or establishes a universal capability ranking.
Bloomberg also reported concerns from anonymous people with direct access about uneven performance on some coding tasks, including front-end design. Google disputed characterizing Argon as underperforming in coding, and another employee described broad internal agreement that it remained frontier-class. With no published task set, sample size or controlled comparison, those accounts cannot be weighed directly against the benchmarks.
Our read is that the restricted rollout gives defenders a useful opportunity to test Argon where success could protect people’s information. The next consequential evidence is whether its discoveries survive expert validation and become deployed fixes. A documented repair to the hospital software would tell readers more about that promise than another benchmark lead.