OpenAI ships a zero-day hunter, then pledges $1B to defense

GPT-6 Astra, launched 3 Sept, is the first OpenAI model to cross the Critical cybersecurity threshold in the company's own Preparedness Framework. Tested without production safeguards, it scored 100% on ExploitBench — up from 78.5% for GPT-5.6 Sol. Translation: with tools and access, it can find previously unknown flaws in well-defended systems and build exploits for them without a human steering each step.

OpenAI shipped it anyway, with the sharp edge gated. Astra went first to business testers in the Daybreak program before rolling out to Plus, Pro, Business and Enterprise tiers plus the API and AWS, and enterprise admins have to switch the cyber capabilities on by hand — off by default at launch.

The counterweight landed a day later: $1 billion in subsidized Daybreak access for people defending things that break loudly — state and local government, water and power utilities, community banks, nonprofits, open-source maintainers. It's discounted access, training and support rather than a grant, aimed to be consumed inside roughly six months, US organizations first. What nobody has published is the price after the subsidy expires.

Why it matters: OpenAI now sells both the lockpick and the lock, and only one of them has a stated price.