AI Risk Clock
Doomsday Clock5 min to midnight
→
🌑 Dark edit
theverge.com2026-10-28

Google launches Gemini 4 Carbon with a ‘significant leap’ in cybersecurity - The Verge

ModelsSafety

Google has released Gemini 4 Carbon, a frontier model whose headline achievement is topping two cybersecurity leaderboards — SEC-bench Pro and ExploitGym — which is the sort of pair of names you get when the branding meeting happens after the third coffee. There is no product page, no API key, no pricing tier — just a press release and a velvet rope. The model is available only to 'a select group of cybersecurity defense companies,' which is corporate for: we've handed the sharpest knife in the drawer to whoever signed the right paper. Everyone else can admire it through the brochure.

The Verge reports 'significant gains' in coding, agentic tasks and cybersecurity, and quotes Google's internal teams having watched Carbon 'run for days to persist through dead ends on complex, multi-step engineering tasks.' Days of persisting through dead ends. That is not a capability, it is a description of every infrastructure team in Britain at 4pm on a Friday, and nobody is naming a benchmark ExploitGym after us. When this eventually reaches 'other partners in its Fairwind Program' — Fairwind, a name chosen by someone who has never been on a boat in poor weather — the rest of us will be reassured that the word 'defense' appears in the partner list.

Meanwhile the genuinely open version of this story we covered a fortnight ago involved a Chinese model with free weights doing complex cyberattacks in controlled tests, which came with a download link and an ethics disclaimer nobody read. Google's answer is a gated model, two leaderboard screenshots, and a programme with a nautical name. The capability is real; the governance is a guest list, and the guest list is written by the people selling the security — which is at least a policy, in the way a bouncer is a policy.

Read this story in another voice
● REC · 2026