Something rare happened this week in cybersecurity. The three biggest AI labs in the world announced models built specifically to protect networks, while security researchers quietly showed how the same technology can be used to break into them. Both stories are true. Both matter to your business.
The Good News: Defenders Just Got Smarter Weapons
Google unveiled Gemini 3.8 Flash Cyber, which it calls its most capable cybersecurity model, and made it available through a new initiative called the Fairwind Program. That program gives priority defenders, including hospitals, telecom providers and government teams, early access to advanced models that help them spot threats faster.
OpenAI followed days later with GPT-6 Astra, which reached the Critical threshold on the company's own cyber safety scale after scoring a perfect 100 percent on ExploitBench, a benchmark for exploit research. In the same breath, OpenAI began blocking requests for exploit demonstration code. Anthropic also shared its own cyber focused work with defenders.
Translation for the rest of us: the people protecting hospitals, banks and power grids are finally getting AI teammates that think in security. That is genuinely good news.
The Other Side Noticed Too
Here is the part that keeps security professionals up at night. The Aurora ransomware group has been using the Cursor AI coding assistant to break into target networks, according to researchers at CloudSEK and Gambit Security. Forescout researchers used Anthropic's Claude to move an exploit from one industrial controller to another, showing how quickly AI can adapt an attack to new hardware. A Russia aligned group is now hiding hostile prompts inside malware to confuse AI powered analysis tools. And Manifold Security found that a poisoned git config file can trick Claude, Codex or Cursor into running attacker code on a developer's machine.
The tools are the same. The intentions are not.
What This Really Means for You
AI did not invent the cyber arms race. It just made both sides faster. And that means the winning move is still the human one.
CISA just published an advisory based on two real red team assessments. In one organization, the simulated attackers roamed freely across workstations and the domain for weeks, completely undetected. In the other, the security team spotted the first sign of intrusion and shut it down within the hour. The difference was not fancier software. It was strong baselines, honest monitoring, and teams that actually talk to each other.
Four Gentle Habits That Still Matter Most
- Patch quickly. CISA added several actively exploited flaws to its catalog this week, and zero day attacks hit VPN appliances, e commerce platforms and routers within days of disclosure. Treat vendor updates like a meeting with your bank.
- Turn on multi factor authentication everywhere it counts. Most recent intrusions began with a stolen password or a credential bypass.
- Watch how your team uses AI. If developers use AI coding agents, keep them on approved repositories and limit what they can access. Treat AI tools like new employees. Least privilege, clear rules, and logs.
- Practice your response. The team that caught the red team did not have better tools. They had rehearsed instincts and a culture where raising the alarm is welcomed.
The news this week can feel like a race between two very fast runners. But remember, one of them is on your team. Keep your systems patched, keep your people curious, and keep your security basics strong. That is still the best strategy, with or without the robots.
Dig deeper: Google, Anthropic and OpenAI unveil cyber AI models · GPT-6 Astra reaches the Critical threshold · Aurora ransomware operators use Cursor AI · CISA red team advisory