AA Index v5.0 登顶,网络安全基准领先
Core Highlights
Google CEO Sundar Pichai announced on X that DeepMind's new frontier model, Gemini 4 Carbon, has officially arrived — Google's second rapid push of the technical frontier after Gemini 4 Argon two weeks earlier. Carbon is described as "the strongest model yet," targeting two directions: state-of-the-art cybersecurity capability and sustained agentic workflows. Internal teams have observed it running continuously for days, pushing through dead ends to handle complex multi-step engineering tasks across large codebases and low-level systems, such as root-cause localization and patching.
What It Does
On unguarded external benchmarks, Carbon tops the Artificial Analysis Intelligence Index v5.0 (score 59, first place); SEC-bench Pro discovers real vulnerabilities at 88.2%, ExploitGym validates exploitability at 54.8%, and CWE-bench produces root-cause patches at 71.0% — all SOTA. Gray Swan's indirect prompt-injection benchmark shows an attack success rate of 0.0%. These figures paint a model that is not only clever but also unusually resistant to being manipulated, a property that matters enormously once it is allowed to act autonomously.
Technical Details
Alignment also improved markedly: unsupported claims are down about 2x versus Argon, high-severity misalignment alerts fall nearly 80%, and Terminal-Bench 4.0 cheating rate drops to 0.0%. Pricing uses an introductory rate of $2 per million input tokens and $10 per million output tokens, doubling after the intro period. The pricing is clearly engineered to seed adoption quickly and then monetize the dependence that follows, a pattern now standard across frontier releases.
How It Compares
In Chrome team tests, Carbon's ability to produce high-quality patches is judged over 10x better than "a large frontier model from another lab." It takes first place on multiple security benchmarks and the composite intelligence index, implying industry-leading position. The subtext is a direct challenge to rival labs: Google is arguing that the frontier is now defined less by chat fluency and more by autonomous, secure, long-running engineering.
Industry Impact and Use Cases
For security teams and heavy coding scenarios, Carbon turns "discover-verify-remediate" into an end-to-end capability — a meaningful increment for defenders. The phased, cautious rollout also shows that the double-edged nature of frontier models on both offensive and defensive sides is being taken seriously, with staged access rather than a sudden public free-for-all.
Practical Takeaways
For security and engineering leaders, the takeaway is that autonomous, long-running coding agents are crossing from lab curiosity to operational tool, and the differentiator is endurance plus resistance to manipulation rather than raw cleverness. Adopt it the same way you would a junior engineer with high autonomy: scoped permissions, mandatory review gates, and monitoring for prompt-injection attempts, since a model that runs for days is also exposed to days' worth of attack surface. The near-zero cheating and injection rates are the metrics that justify that trust, so validate them on your own red-team prompts before widening access. On cost, treat the intro pricing as temporary and model the post-intro tier into any build-versus-buy decision. Strategically, Google is signaling that the frontier is now defined by secure, sustained agency — a bar that reframes the competitive race away from chat and toward trusted autonomous work.
Wider Context
The strategic backdrop is Google's attempt to reframe the frontier around trusted autonomy rather than conversational prowess, a move that suits a company with deep enterprise and developer reach. If "secure, long-running agent" becomes the headline metric, Google's heavy investment in alignment and safety benchmarking pays off as differentiation, not just compliance. For competitors, the pressure is to match not only capability but the safety properties buyers now demand before granting autonomy. And for the wider public, the arrival of models that can patch systems for days at a time is a double-edged milestone: enormous productivity upside, but also a new class of risk that requires the same staged, monitored adoption the lab itself is modeling. The story of Carbon is less "a smarter model" than "a more trustworthy autonomous one."
Bottom Line
Treat Carbon like a high-autonomy junior engineer: scoped permissions, mandatory review, and red-team monitoring for prompt injection. Google is competing on trusted autonomy, and that is the metric buyers should evaluate before granting any model long-running access.
The era of "release it and see" is giving way to "release it carefully and watch," and Carbon is Google's argument that trusted autonomy is the new frontier.