Google unveils Gemini 4 Carbon with security benchmark claims

Google on Wednesday 29 October 2026 announced Gemini 4 Carbon, a new frontier model that the company says has state-of-the-art cybersecurity and long-horizon capabilities, according to a report by The Hacker News.

Google said the model shows significant gains across coding and agentic tasks, as well as across capabilities it described as critical for cybersecurity defence. The company cited a score of 88.2% on SEC-bench Pro, a benchmark for finding vulnerabilities, and 54.8% on ExploitGym, a benchmark for exploitation. Google said these results give it a firm lead on both tests.

The announcement is notable for anyone running AI agents in security workflows, because the same abilities that let a model find and patch a flaw also let it find and exploit one. A model that scores well on an exploitation benchmark is, by definition, a model that can write working attack code against test targets. Google's framing is defensive, but the benchmarks it chose measure both halves of the job.

Several caveats are worth stating plainly. The scores reported so far come from the vendor's own announcement, as relayed by The Hacker News. There is no independent reproduction of the figures in the source material. Benchmark numbers for agentic security tasks are also sensitive to how the harness is built: what tools the agent is given, how many attempts it gets, and how a task is scored as passed or failed. None of those details are available from the material published so far.

It is also unclear from the announcement what guardrails, if any, Google has placed around the exploitation capabilities, or who gets access to them. The source does not say whether the model is generally available, restricted to approved customers, or limited to internal use. Nor does it describe any safety evaluation or release policy attached to the cybersecurity results.

"Long-horizon" capability, the other claim in the announcement, refers to a model's ability to keep working on a task across many steps without losing track of the goal. For security work this matters a great deal. Finding a vulnerability in a large codebase is rarely a single prompt. It usually involves reading code, forming a hypothesis, testing it, and backtracking when the hypothesis is wrong. Models that drift or forget earlier steps perform poorly at this even when they score well on single-shot coding tests. Google said Gemini 4 Carbon improves here, but the source does not include a specific figure for long-horizon performance.

For defenders, the practical question is not the benchmark number but whether an agent of this kind can be trusted inside a code review or patching pipeline. An agent that proposes patches needs review, because a confident but wrong patch can introduce a new flaw while appearing to close an old one. Benchmarks that measure patch success rates do not usually measure how often a model introduces a regression.

The Hacker News report is the only account of the announcement available at the time of writing, and its summary of Google's claims is the basis for the figures above. Independent evaluation of the model's security performance has not yet been published.