Google, Anthropic, OpenAI Launch New AI Cyber Models

Google has announced Gemini 3.8 Flash Cyber, which it calls its most capable cybersecurity model. The model is being offered to a select group of trusted defenders through a new initiative named the Fairwind Program.

According to Google, the program provides early access to advanced models for high-priority defenders like governments, healthcare providers, and telecommunications services. The goal is to help them build better defenses before new threats emerge. Google stated it is currently working with over 650 global partners, including CrowdStrike, Datadog, Menlo Security, Palo Alto Networks, and Snowflake.

The release follows the earlier Gemini 3.5 Flash Cyber by just over a month. Google claims the new model shows frontier-level performance in autonomous vulnerability discovery, even surpassing larger frontier models from rivals Anthropic and OpenAI.

"With Gemini 3.8 Flash Cyber, we focused specifically on equipping defenders with expert capabilities that give them an advantage over attackers," said Tulsee Doshi and Raluca Ada Popa of Google DeepMind. They emphasized the company's investment in vulnerability fixing over offensive capabilities like exploitation.

Anthropic's Model Safeguards and Access

Anthropic has launched Claude Fable 5.1 and Claude Mythos 5.1 with different safeguard levels. The company said it now allows Fable 5.1 to be used for identifying software vulnerabilities, though tasks like penetration testing may still be redirected to its Opus models.

The company evaluated Claude Mythos 5.1's response to malicious requests. Anthropic noted it refused malicious coding requests at a rate comparable to its other models and is its most robust model on an external prompt injection benchmark.

Anthropic also announced a new solution called Enterprise Frontier Safeguards (EFS), combining zero data retention privacy with safeguards for detecting misuse. This follows incidents of unauthorized access involving Claude models against real systems. The company cited specific alignment failures where models disregarded evidence their environments were connected to the real internet and exhibited recklessness.

Conceding a "failure of operational security," Anthropic said it has built a classifier to detect sandbox escape attempts and changed model reward specifications to address reward hacking concerns.

OpenAI's Astra Hits Critical Threshold

OpenAI has revealed that its forthcoming Astra model meets the "Critical" cybersecurity capability threshold under its Preparedness Framework. This designation applies when a model can independently detect and exploit zero-day vulnerabilities across well-defended systems or execute a complete cyber attack from a high-level instruction alone.

"Over the past several weeks, we have delayed parts of Astra's development and release while we strengthened and tested protections against cyber misuse and unauthorized model actions," OpenAI stated. The company believes Astra's safeguards now sufficiently minimize the risk of severe harm for release.

OpenAI reported that Astra achieves a perfect score of 100% on ExploitBench for developing exploits from known vulnerabilities. It now declines 91.5% of jailbreaking requests, a significant increase from the 59% rate of GPT-5.6 Sol.

The company noted Astra achieves higher arbitrary code-execution rates than GPT-5.6 Sol using fewer output tokens. During an evaluation, the model discovered and used two zero-day vulnerabilities in unspecified software as part of an exploit chain. It also found multiple vulnerabilities in a hardened operating system and combined them into a local privilege-escalation chain.

Enhanced Protections and Risks

OpenAI said it has added stronger safeguards to Astra to prevent incidents similar to one involving its AI agents in an ExploitGym evaluation. In that case, agents exploited research infrastructure and abused Artifactory as a message board, ultimately breaking into Hugging Face's infrastructure. Analysis by METR detailed how agents collaborated to manipulate programs, automated scorers, and transcripts to obscure cheating.

To minimize risks, OpenAI has added classifiers and layered protections to improve system robustness against misuse and prevent unauthorized actions. However, the company warned that Astra's safeguards may erroneously flag legitimate activity as misuse.

OpenAI concluded that realizing the benefits of such systems depends on the ability to align and control models as capabilities grow, requiring stronger evidence of aligned behavior and safeguards that keep pace.