Google's AI 'Carbon' Targets Software Flaws

Google's newest AI model, Gemini 4 Carbon, is positioned as a significant advancement in cybersecurity and software engineering. This powerful AI is engineered to detect software vulnerabilities, assess their exploitability, and assist in developing patches to fix them, potentially accelerating threat response for security teams.

According to Google, Carbon demonstrated strong performance on industry benchmarks, achieving an 88.2% score on SEC-bench Pro and 71% on CWE-bench for vulnerability identification and patching, respectively. The company also reported that Carbon generated secure code at more than double the success rate of its predecessor, Gemini 4 Argon.

Elon Musk notably praised the AI's capabilities with a single word, "Impressive," on X. Google is implementing a cautious, phased rollout for Gemini 4 Carbon, beginning with cybersecurity defenders via its 'Fairwind Program.' This limited release allows for further testing and refinement of safeguards before wider distribution, reflecting industry-wide challenges in balancing advanced AI capabilities with security concerns. Pricing is set at $2 per million input tokens and $10 per million output tokens.