Artificial Intelligence
Google rolls out Gemini 4 Argon
Google said it was widening access to Gemini 4 Argon and its Fairwind Program, while describing safety tests and controls aimed at limiting misuse, misalignment and prompt injection.
- Google said it had begun a broader rollout of Gemini 4 Argon and would expand its Fairwind Program to more trusted cyber defenders.
- Google said Wiz had used Argon to find a critical bug in hospital software that earlier models had missed, helping protect hundreds of thousands of patient records.
- Google said its own security teams had used Argon to find and fix bugs in internal and open-source software, including vulnerabilities handled through responsible disclosure.
- Google said Argon solved all 12 simulated end-to-end cyberattack scenarios in its testing, up from five in the previous generation, and that prompt-injection attack success fell to 0.7% on Gray Swan’s benchmark.
Google said it was starting a wider release of Gemini 4 Argon and was extending its Fairwind Program to more trusted cyber defenders. It described the model as useful for coding, finance, legal work, cybersecurity and science.
During a limited-access phase, Google said Wiz used the model to uncover a critical flaw in hospital software that earlier systems had not found. Google also said its own security teams used Argon to identify and patch bugs in internal and open-source software, including issues now going through responsible disclosure.
The company said its safety work focused on misuse, misalignment and prompt injection. It said the model did better than earlier versions in cyber and biology tests, and that it was more likely to refuse harmful requests and resist jailbreak attempts.
Google said it had found no covert behaviour in its honeypot tests for misalignment. It also said an automated monitor was used in products, internal coding agents, training runs and evaluations, and that the model scored 0.7% attack success on Gray Swan’s indirect prompt-injection benchmark.
Named in this story
Companies
- published the article and said it was widening access to the model
- Wiz
- used Argon to uncover a critical bug in hospital software
- Gray Swan
- provided the prompt-injection benchmark used in Google’s testing
Organisations
- Fairwind Program
- the access programme being expanded to additional defenders
Products and systems
- Gemini 4 Argon
- the model being rolled out and assessed
How the source tells it
A promotional safety brief that mixed technical detail with self-praise and reassurance.
- vendor boosterism superlative performance claims and first-person praise for its own model
- reassurance repeated emphasis on guardrails, monitoring and defence-in-depth controls
- threat framing high-risk misuse scenarios involving cyberattacks and biological threats