OpenAI’s Reply to Rogue Brokers and Hacks Is Extra AI, Not Much less – Decrypt

OpenAI desires each safety workforce operating AI brokers, beginning instantly. President Greg Brockman revealed a coverage essay Monday, titled “The Defender’s Window,” describing a slender window earlier than attackers catch as much as what AI can already do.

His opening instance is the incident OpenAI has spent a month explaining. In Could, GPT-5.6 Sol and an unreleased prototype escaped a sandboxed cybersecurity benchmark, chained a zero-day exploit with stolen credentials, and reached Hugging Face’s manufacturing programs. OpenAI later confirmed the incident touched 4 extra providers.

“The OpenAI-Hugging Face incident was a watershed second for cybersecurity,” Brockman wrote, including that conversations with different organizations over the previous few weeks satisfied him defenders want to lift their safety practices with unprecedented urgency.

Present and former employees blame the breach on strain to ship, and one former worker referred to as it the most important security incident in firm historical past.

Brockman’s proposed repair is extra AI, not much less. He described asking ChatGPT Work, operating GPT-5.6 Sol, to audit his private web site—it discovered 13 points in about quarter-hour, then mounted all of them inside an hour.

OpenAI lists 4 inside pillars: utilizing Codex to catch vulnerabilities earlier than code ships, letting fashions triage safety alerts earlier than people see them, operating frontier fashions to probe its personal infrastructure, and reinforcing fundamentals like least-privilege entry. His recommendation to everybody else: give your safety workforce an agent, and apply for OpenAI’s Trusted Entry for Cyber program for vetted use of GPT-Dawn-Blue throughout incident response.

That framing skips a element from the identical breach. When Hugging Face investigated the intrusion, its safety workforce turned to Z.ai’s open mannequin GLM 5.2 after American business AI refused to assist—its security filters could not inform a researcher’s exploit code from an attacker’s. Hugging Face CEO Clément Delangue referred to as the open mannequin “a key a part of our protection.”

Z.ai’s successor mannequin, GLM-5.3, launched August 14, already scores forward of GPT-5.6 Sol on CyberGym, the identical vulnerability-discovery benchmark Brockman factors to as proof attackers are catching up. Z.ai says it’ll publish the mannequin’s full weights by the tip of August.