GPT-5.6 Security: Hardware Passkeys for Advanced AI Access

OpenAI has released its new GPT-5.6 family of models, including GPT-5.6 Sol, which shows significant performance gains in cybersecurity benchmarks. These new models also consume fewer tokens during operation. Here's the thing: individual members of OpenAI's Trusted Access for Cyber program now face a new security requirement. They must adopt hardware-backed passkeys to maintain access to the most capable models. The deadline for this is September 1, 2026. If they don't comply, they will lose access to these advanced models. What's interesting is GPT-5.6 Sol achieved 73.5% on ExploitBench, compared to 47.9% for its predecessor. It also improved on ExploitGym, from 15.1% to 24.9%, and on SEC-Bench Pro, from 45.8% to 71.2%. OpenAI previously partnered with Yubico for security keys and is now offering preferred pricing for Yubico keys to users without them. The bottom line is this move highlights a growing focus on robust security for advanced AI access.