{"id":14573,"date":"2026-10-31T06:01:19","date_gmt":"2026-10-31T06:01:19","guid":{"rendered":"https:\/\/krikis.lt\/gemini-4-carbon-saugumas\/"},"modified":"2026-10-31T06:01:19","modified_gmt":"2026-10-31T06:01:19","slug":"gemini-4-carbon-saugumas","status":"publish","type":"post","link":"https:\/\/krikis.lt\/en\/gemini-4-carbon-safety\/","title":{"rendered":"Gemini 4 Carbon \u2013 a new step in security technology"},"content":{"rendered":"

\u201eGemini 4 Carbon is Google DeepMind's latest artificial intelligence model, which the company introduced on October 28, 2026. According to PCMag.com, the model is designed for continuous agent-based workflows and provides "a significant leap in cybersecurity capabilities." Testing shows that Carbon outperforms not only the previous Gemini 4 Argon model, but also competitors such as OpenAI GPT-6 Astra and Anthropic Claude Opus 5.5 in a variety of benchmarks, from coding to scientific computing.<\/p>\n

Major progress in cybersecurity<\/h2>\n
\"Server<\/figure>\n

According to PCMag.com, the Gemini 4 Carbon is able to detect critical vulnerabilities, validate them, and generate high-quality patches. The model also has exceptional resistance to \u201eprompt injection\u201c attacks, a technique in which cybercriminals try to force AI to bypass its own protection mechanisms. Internal tests showed a success rate of 0.0 % in this area, compared to 0.7 % for the Gemini 4 Argon and 1.0 % for the Claude Opus 5.5.<\/p>\n

Evaluation indices and differences in results<\/h2>\n
\"Large<\/figure>\n

Crypto Briefing reports that Gemini 4 Carbon secured first place in the Artificial Analysis Intelligence Index v5 with a score of 59. This is a 9-point jump from its sibling Argon, which scored 50. The model excels in four of the nine components of the index, including HLE-Diamond and AutomationBench-AA. The most important are cybersecurity tests: SEC-bench Pro score of 88.2 % (vulnerability detection) and CWE-bench 71.0 % (code patch generation). In addition, the ExploitGym test gives a score of 54.8 %, and GraphWalks \u2013 96.2 % for long-term context understanding.<\/p>\n

Access, pricing and distribution strategy<\/h2>\n

Both sources indicate that Gemini 4 Carbon will initially be available in a limited capacity to select cybersecurity professionals through Google\u2019s Fairwind program. PCMag.com reports that the initial API pricing model is $2 per million input tokens and $10 per million output tokens. Crypto Briefing reports that the standard price is $4 per million input tokens and $20 per million output tokens, but there is a 50% discount for the initial phase. Pricing information for local markets (UAE, Saudi Arabia, Egypt) has not yet been announced, so users are advised to wait for an official regional announcement.<\/p>\n

Competitive environment and future prospects<\/h2>\n

\u201eGemini 4 Carbon highlights Google\u2019s approach to security AI, initially focusing on the needs of security teams rather than the general public. This reflects concerns that models with such power could be misused. The guidelines Google has in place allow it to control who gets access first, while maintaining a competitive advantage over OpenAI and Anthropic.<\/p>\n

Technical challenges and practical application<\/h2>\n

While Gemini 4 Carbon shows impressive results, its practical application in the business environment is just beginning. The model\u2019s ability to generate reliable patches can help organizations reduce the number of unpatched vulnerabilities, but this requires integration into existing security platforms and processes. In addition, prompt injection resilience provides an additional layer of security when AI is used as an aid to automated tasks.<\/p>\n

Conclusions<\/h2>\n

\u201eGemini 4 Carbon is a significant step forward in cybersecurity technology, as evidenced by data from both PCMag.com and Crypto Briefing. The model\u2019s high benchmark scores, exceptional attack resistance, and strategic access management demonstrate that Google is not only pursuing technological superiority, but also responsible AI deployment. While access is currently limited, we can expect a broader release of the model in the future, which could change the way organizations approach automated vulnerability detection and patch generation.<\/p>\n

Sources<\/h2>\n