The three major laboratories have unveiled their cyber-defence models in quick succession, with restricted access programmes to prevent the same capabilities from being used for attack.
Google presents Gemini 3.8 Flash Cyber as its most advanced model on security tasks, and accompanies it with a program called Fairwind that reserves early access to defenders: public administrations, health institutions and specialized publishers. More than 650 partners are already in the world, including CrowdStrike, Datadog and Palo Alto Networks.
Manoeuvring responds to a problem that has become central: a model capable of identifying a unique fault alone is just as capable of exploiting it. Anthropic has therefore limited Mythos 5.1 to trusted access programs covering cybersecurity and life sciences, and deployed an Enterprise Frontier Safeguards device combining zero data retention and detection of misuse.
On the OpenAI side, Astra crossed the critical capacity threshold in cybersecurity: the model was able to identify and exploit zero-day vulnerabilities independently, and discovered unknown defects in hard operating systems. It reached 100% on ExploitBench for the writing of exploits and refused 91.5% of attempts to circumvent its safeguards, compared with 59% for GPT-5.6 Sol. The broadcast is through Daybreak Blue.
Read the full article on the original source
Go to the source →