GPT-6 CYBER: ADVANCING THE FRONTIER OF ALIGNED CYBERSECURITY AGENTS

Summary from the TLDR InfoSec newsletter

OpenAI's GPT-6 Cyber built working exploits for 948 Linux kernel vulnerabilities, more than five times GPT-5.6 Sol's count with the same harness, and found 81 previously unknown high-severity Firefox bugs. OpenAI also trained it to stay within the scope a researcher sets: on a honeypot evaluation, it avoided easy-to-exploit, out-of-scope services in 99.93% of runs without a safety prompt and 100% with one, enough for OpenAI to rate it Critical for cybersecurity under its Preparedness Framework. Approved defenders reach it through Daybreak Red, which requires IP allowlisting by default and runs misuse monitoring, and a new tier gives critical infrastructure teams authorized OT/ICS testing.