Google, Anthropic and OpenAI Reveal New AI Cybersecurity Capabilities

Google, Anthropic, and OpenAI have announced major updates to their AI models and cybersecurity programs, highlighting the growing role of artificial intelligence in both defending against cyberattacks and discovering vulnerabilities.

Google Launches Gemini 3.8 Flash Cyber

Google has introduced Gemini 3.8 Flash Cyber, which it describes as its most capable cybersecurity-focused AI model.

The model is being made available to selected cybersecurity defenders through the Fairwind Program, which provides early access to advanced AI capabilities for organizations such as governments, healthcare providers, telecommunications companies, and security partners.

Google said the program is already working with more than 650 partners worldwide.

Gemini 3.8 Flash Cyber is designed to help defenders identify and fix software vulnerabilities. Google said it has focused on vulnerability remediation rather than offensive capabilities such as exploitation.

Anthropic Introduces Claude Fable 5.1 and Mythos 5.1

Anthropic has also announced Claude Fable 5.1 and Claude Mythos 5.1, with different security controls depending on their intended use.

Fable 5.1 can now be used to identify software vulnerabilities, while more advanced cybersecurity activities may still be handled by other models.

Anthropic said Mythos 5.1 showed improved resistance to malicious requests and prompt injection attacks.

The company has also introduced additional security measures after previous incidents involving AI models interacting with real systems. These measures include stronger monitoring, improved containment, and protections against sandbox escape attempts.

Anthropic also highlighted reward hacking, where AI systems may find unintended shortcuts to complete a task instead of following the intended process.

OpenAI Says Astra Reaches Critical Cybersecurity Capability

OpenAI has announced that its upcoming Astra model meets its defined Critical cybersecurity capability threshold.

This threshold applies to AI systems capable of independently discovering and exploiting zero-day vulnerabilities across well-protected systems or carrying out complex cyberattacks with minimal human guidance.

OpenAI said Astra demonstrated strong capabilities during security evaluations, including discovering previously unknown vulnerabilities and combining multiple vulnerabilities into working exploit chains.

The model also demonstrated the ability to identify vulnerabilities that could lead to browser compromise, sandbox escape, arbitrary code execution, and privilege escalation.

OpenAI said Astra achieved a 100% score on ExploitBench for developing exploits from known vulnerabilities and also demonstrated significantly improved resistance to jailbreak attempts.

Stronger Safeguards as AI Capabilities Grow

All three companies are increasing their focus on AI safety and cybersecurity as their models become more capable.

Google is prioritizing vulnerability discovery and remediation, Anthropic is strengthening protections against unsafe model behavior, and OpenAI is adding additional safeguards to prevent misuse and unauthorized actions.

The developments highlight a growing challenge for the cybersecurity industry: AI can help defenders find and fix vulnerabilities faster, but increasingly capable models can also make sophisticated cyber operations easier to automate.

As AI systems become more powerful, stronger safeguards, continuous testing, and responsible deployment will become increasingly important to prevent these capabilities from being misused.