AI Helped Hackers Break Into OpenAI in Hours for $3,000

Three security researchers from a firm called Hacktron used Claude Opus 5 — an AI built by OpenAI's rival, Anthropic — to break into OpenAI's private code storage. That storage is a "monorepo," a single giant folder where a company keeps nearly all of its software in one place. They reportedly got in within hours of the AI's public release, and the entire operation cost less than $3,000 in computing credits. That's cheaper than a decent laptop.

The AI didn't magically guess a password. It chained exploits, meaning it found several small weaknesses and strung them together, the way a burglar checks a window, then a side door, then a loose lock. That kind of work used to take a skilled human team weeks of careful digging. Now it takes hours and pocket change. The hackers call themselves "white hat," meaning they break in to expose problems rather than to steal. Still, there's no guarantee the next person with $3,000 is friendly.

What matters is what they didn't get. OpenAI's model weights — the AI's actual learned "brain," the crown jewels — were reportedly not part of the breach. Reading a company's code is not the same as copying its AI. But internal code and tools still reveal what a company is building, how it works, and where it's fragile.

Most of us don't run AI companies, but we all hand our data to someone. If $3,000 of AI can crack a top-tier tech company's defenses in an afternoon, smaller targets — your bank, your hospital, your employer — are far more exposed. Expect a rush of spending on AI-powered defense, and expect companies to start asking uncomfortable questions about what their own AI tools might accidentally reveal.