Google, Anthropic and OpenAI Roll Out Cyber-Focused AI Models and Access Programs
Three of the biggest AI labs are pitching new tools at defenders, with early access for hospitals, governments and telecoms.
Key points
- Google announced Gemini 3.8 Flash Cyber on Wednesday, calling it its most capable cybersecurity model to date.
- Access is gated through a new Fairwind Program aimed at governments, healthcare providers and telecoms.
- Anthropic and OpenAI launched parallel efforts the same week to put security-tuned models in defenders' hands.
- None of these models replace existing security tools or the analysts who use them.
Google has a new AI model built specifically for cybersecurity work, and it isn't letting just anyone use it yet.
The model is called Gemini 3.8 Flash Cyber. Google described it on Wednesday as its most capable cybersecurity model: an AI system tuned to help defenders spot attacks and understand malicious code faster than a human analyst working alone.
Access is gated through something Google is calling the Fairwind Program, a members-only preview where high-priority defenders get the newest model before it opens more broadly.
Anthropic and OpenAI announced their own security-focused releases the same week. The pitch across all three is roughly the same: defenders are outnumbered, and a well-tuned AI model can read logs, sift alerts faster than any solo analyst can manage.
Who actually gets to use these tools?
Not the general public. Google's Fairwind Program targets organisations it considers high-priority defenders: government agencies, healthcare providers and telecommunications companies that keep phone and internet networks running.
Those are the sectors hit hardest when a cyberattack lands. Ransomware on a hospital (malicious software that locks files until a ransom is paid) can delay surgeries. An attack on a telecoms firm can take critical communications offline. Google's argument is that these defenders should get the sharpest tools first.
Anthropic and OpenAI are structuring their programs similarly, with vetted access for security teams rather than open sign-ups. The question of whether these same labs' own products can be trusted is one we've been tracking: our 2 September story found security experts debating whether AI safety limits slow defenders more than attackers.
What can a "cybersecurity model" actually do?
It reads security data at speed and explains what it finds. That's the honest version. These models are trained on more security-relevant material than a general chatbot, so they're better at recognising the fingerprints of a phishing email (a fake message designed to steal credentials) or malware buried in a log file.
Google says Gemini 3.8 Flash Cyber helps analysts triage alerts, reverse-engineer suspicious files and write up incidents. Anthropic and OpenAI make similar claims.
Be sober about this. A security-tuned model is still a language model. It will sometimes get things wrong, and no vendor is suggesting it replaces the human analyst reading the screen at 3am. As reported by The Hacker News, all three labs are framing these as defender assistants, not replacements.
| Lab | Offering | Early access route |
|---|---|---|
| Gemini 3.8 Flash Cyber | Fairwind Program | |
| Anthropic | Security-focused model release | Vetted defender program |
| OpenAI | Cyber AI model and safeguards | Vetted defender program |
Why now?
Because attackers are already using AI. Criminal groups have been caught using general-purpose chatbots to write phishing emails and accelerate basic coding tasks. If defenders don't get purpose-built tools, they'll fight the next few years with a handicap.
There's also a competitive angle. Microsoft has been selling its Security Copilot product to enterprise customers for more than a year. Google, Anthropic and OpenAI are staking out ground in the same market, with the twist that early access goes to critical-infrastructure defenders rather than the highest bidder.
What should ordinary people take from this?
Not much, directly. You won't download Gemini 3.8 Flash Cyber. But the hospital that stores your records and the phone company that carries your calls may soon have sharper tools for spotting intruders. That's the point.
Whether it works in practice is the real question. Given that we've also reported on security flaws in the software wrappers around these same AI systems, the next twelve months will be instructive.
