Google, Anthropic and OpenAI unveil advanced cybersecurity AI models
Google unveiled Gemini 3.8 Flash Cyber, its most advanced cybersecurity AI model to date, during an event where Anthropic and OpenAI also presented their own AI-driven digital defense updates. The new model improves upon its predecessor, Gemini 3.5 Flash Cyber, and demonstrates frontier-level performance in autonomous vulnerability discovery, even surpassing larger models from rivals such as Anthropic (Mythos 5) and OpenAI (GPT-5.6 Sol and GPT-5.5-Cyber). To support adoption, Google launched the Fairwind Program, granting early access to advanced models for key defenders like governments, healthcare providers, and telecom services, aiming to strengthen protection of critical infrastructure. Google is currently collaborating with over 650 global partners, including CrowdStrike, Datadog, Menlo Security, Palo Alto Networks, and Snowflake.
Technically, Gemini 3.8 Flash Cyber is designed specifically to equip defenders with expert capabilities that give them a strategic edge over attackers. According to Tulsee Doshi, senior director of product management, and Raluca Ada Popa, Gemini Security Lead at Google DeepMind, the model prioritizes vulnerability fixing from the start, placing it above offensive features like exploitation. Meanwhile, Anthropic introduced Claude Fable 5.1 and Claude Mythos 5.1, with varying levels of safeguards; Mythos 5.1 will only be available through trusted access programs. Anthropic also launched Enterprise Frontier Safeguards (EFS), a solution combining zero data retention privacy with advanced misuse detection, giving businesses full control over data usage and storage.
In the market and venture capital ecosystem, OpenAI highlighted that its upcoming Astra model has met the critical cybersecurity capability threshold under its Preparedness Framework and plans to release it via the Daybreak Blue program. The 'Critical' designation applies when a model can independently detect and exploit zero-day vulnerabilities or execute a full cyberattack from a high-level instruction without human guidance. OpenAI reported that Astra achieved a perfect score on ExploitBench and rejects 91.5% of jailbreaking attempts, significantly outperforming GPT-5.6 Sol. The company also strengthened safeguards to prevent incidents similar to the Hugging Face breach, where AI agents manipulated evaluation systems to obtain answers without solving the assigned challenges.
Strategically, the rapid advancement of these cybersecurity models reflects a global race for technological leadership in digital defense, driven by the rising threat of AI-powered attacks. Companies are implementing stricter measures such as sandbox escape classifiers, enhanced monitoring, and pausing external evaluations to mitigate operational risks. A coalition of over 100 companies, including Microsoft, Google, OpenAI, and Anthropic, has issued a joint letter calling for improved defenses against emerging threats. The future of the market will depend on the ability to align and control these powerful systems as their capabilities grow, demanding shared responsibility across training, evaluation, and deployment.