🚨 GPT-5.6 just made a massive jump in offensive cybersecurity tests

OpenAI’s newest model showed major gains over GPT-5.5:

• ExploitBench 2 increased from 47.9% to 73.5% • ExploitGym performance rose from 15.1% to 24.9% • Exploit Gym also reached 33.7% given 6 hours to work • Uses fewer tokens than previous models on complex security tasks • OpenAI says it still failed to autonomously complete full exploit chains

The cybersecurity gains are incredible.

Are frontier models becoming too capable to release without restrictions?