Gemini 4 Carbon Puts AI on the Cyber Front Line
Google DeepMind's latest frontier model targets the full vulnerability lifecycle—detection, exploitation, and patching—before most users can access it.
By The ai-compute Desk
Two weeks after shipping Gemini 4 Argon, Google DeepMind is back with Gemini 4 Carbon—a model built explicitly around the cybersecurity defense lifecycle. It tops the Artificial Analysis Intelligence Index v5.0 with a score of 59, and its benchmark numbers across vulnerability finding, exploitation, and patching are the strongest the lab has published. [1]
The headline figures are hard to ignore. Carbon scores 88.2% on SEC-bench Pro for finding vulnerabilities in real-world software, 54.8% on ExploitGym for validating which flaws are actually exploitable, and 71.0% on Collinear's CWE-bench for patching at the root cause—all evaluated without guardrails. [1]
Carbon is also built for the kind of sustained, multi-step work that security engineering demands. Internally, Google's teams have seen the model run for days to persist through dead ends to tackle complex, multi-step engineering tasks like root-causing and patching subtle bugs across large codebases and low-level systems. [1] On WebDev Arena it jumped +89 Elo points to land at rank four with a score of 1769, signaling broader engineering gains beyond pure security. [1]
Raw capability is only part of the story. What makes Carbon worth watching is the parallel effort to constrain its behavior—a bet that a model which won't cheat, fabricate, or overstep is more useful in a security context than one that merely scores well on benchmarks.
On the alignment side, Carbon reduced successful cheating on Terminal-Bench 4.0 to 0.0%, with attempts to cheat dropping to 0.4%. [1] High-severity misalignment alerts in live coding environments dropped by almost 80%, and the model simultaneously reduced overreach by 26% and under-service by 29%. [1]
On prompt injection—a critical attack vector for agentic systems—Carbon achieves 0.0% attack success rate on Gray Swan's Indirect Prompt Injection (IPI) Benchmark, improving on Argon's already-low 0.7%. [1] Cloud security platform Wiz also reports Carbon leading on benchmarks for web penetration testing, a core capability for Wiz's Scan for Good initiative. [1]
Access is intentionally staged. Google is engaged in the U.S. government's voluntary process for pre-release model testing, and Carbon is being rolled out first to a subset of cyber defenders through its Fairwind Program before reaching developers, businesses, and consumers. [1] Pricing mirrors Argon: $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at 95% off the input token price. [1]
The staged rollout reflects a broader tension in AI deployment: the same capabilities that make a model useful to defenders make it dangerous in other hands. Carbon's architecture is a wager that alignment and capability can be trained upward together—not traded off against each other.
Carbon's release schedule—two flagship models in a fortnight—suggests Google DeepMind is running its research pipeline and its product calendar in tighter sync than before. [1] Whether the Fairwind cohort's early feedback accelerates or complicates that cadence will determine how quickly the rest of the market can put it to work. [1]
Carbon reduced successful cheating on Terminal-Bench 4.0 to 0.0%, with attempts to cheat dropping to 0.4% — all while improving overall accuracy.Google DeepMind