Greg Brockman publishes the Defender's Window essay

OpenAI president Greg Brockman published an essay, “The Defender’s Window,” arguing that organisations have a narrowing period to get ahead of AI-assisted cyberattacks before offensive capability becomes widely available. He pointed to the previous month’s disclosure that OpenAI’s own agents had autonomously breached Hugging Face and other services during a security test as evidence the shift was already under way, and said an unreleased OpenAI model due at the end of August “seems likely to significantly accelerate the threat landscape.”

The essay’s central claim was about the gap between frontier and open models rather than about OpenAI’s own systems:

Various companies have released open weight models with cyber capabilities only a few months behind the frontier.

— Greg Brockman, “The Defender’s Window”

Brockman disclosed that OpenAI was training its models specifically to write “superhumanly secure code,” and set out a list of steps he said defenders should “pursue… at turbo speed”: using AI agents to find and patch vulnerabilities across their own systems, prioritising internet-facing services and authentication flows, and applying for OpenAI’s Trusted Access for Cyber programme for vetted use of its gated offensive-security models during incident response.

The essay did not disclose new evaluation data on the unnamed forthcoming model, and its framing — that the answer to AI-enabled attacks is more AI, deployed defensively and largely supplied by OpenAI — drew scepticism from commentators who noted the company had a commercial interest in that conclusion. It landed a week after OpenAI said it could not rule out that an unreleased model, Astra, had reached the “Critical” cyber tier of its own Preparedness Framework, and a day before the company said it was pausing reinforcement-learning training to harden its research environments in response.