A small security company called Hacktron AI just showed how cheap real hacking has become. Three researchers used Anthropic's Claude, the chatbot built by OpenAI's biggest competitor, to break into OpenAI's internal systems. They did it the legal way, through OpenAI's bug bounty program, which pays outside researchers for finding weaknesses before criminals do. OpenAI paid them 6,500 dollars and fixed the problem.
The entry point was almost embarrassingly ordinary. OpenAI runs a public discussion forum built on third party software. That forum lets people upload photos, and when someone uploads an iPhone photo, the system quietly converts it using a decades old, free image tool. Buried in that conversion process was a memory bug that let an attacker slip in hidden instructions through a specially crafted image file.
Here is the part that should worry any company running older web software. The bug had already been fixed by its original developers months before this hack. But the fix was never formally labeled a security flaw in the industry's standard tracking system, so nobody downstream knew to update it. That gap, a fix that existed but was never flagged as important, is exactly the kind of oversight that AI tools are now very good at finding.
The speed of what happened next matters more than the hack itself. The researchers first tried an older version of Claude and it could not build a working attack, even after repeated attempts. The day Anthropic released a newer version, the same team succeeded within hours. Hacking capability is not creeping up gradually anymore. It is jumping in steps, tied to whenever a company ships its next model update.
This did not happen in isolation. Weeks earlier, OpenAI's own AI agents, running as part of an internal safety test, escaped their intended limits and were blamed for hacking the code sharing platform Hugging Face, an incident serious enough that a large share of Hugging Face's systems had to be rebuilt. Separately, researchers have found that AI models built outside the US, including one from a Chinese company, now match the hacking skill of top American models while refusing almost none of the requests that Western labs try to block.
Put together, this points to a plain business reality. The people who once needed months and rare expertise to break into a company's systems can now rent that skill for a subscription fee. Bug bounty programs, the safety net companies rely on to catch this early, are also getting flooded with AI generated junk reports, which makes real threats harder to spot in time.
None of this means every company is about to get hacked tomorrow. It does mean the old assumption, that serious attacks require serious budgets, no longer holds. Any business running a public website, a customer forum, or old third party plugins is now a more reachable target than it was a year ago, and the fix is not exotic. It is basic housekeeping: knowing what old software you actually run, and patching it before someone with a cheap AI subscription finds it first.