Researchers Used Anthropic’s Claude to Break Into OpenAI Employee Accounts. Here’s What Actually Happened.
A security research team called Hacktron AI used Anthropic‘s Claude models to help exploit vulnerabilities tied to OpenAI’s community forum and, in the process, gained access to several OpenAI employees’ ChatGPT accounts. One of those accounts had a Codex connection reaching all the way into OpenAI’s internal GitHub environment. This was authorized research, done under responsible disclosure, and OpenAI has since fixed the issues and revoked the relevant authentication tokens.
The Wall Street Journal broke the story on September 17, 2026, with Forbes and a handful of other outlets adding detail the next day. The actual research happened back on July 25; the public didn’t hear about it until the coordinated disclosure process wrapped up.
What makes this worth paying attention to isn’t just that it happened, but how. Claude wasn’t off doing this on its own; three human researchers were driving the whole thing, choosing the attack path and deciding where to draw the line. But by Hacktron’s account, the newer Claude Opus 5 model took what had been a genuinely hard exploit-development problem and cracked it in a matter of hours.
Where it started: a forum, not ChatGPT itself
The entry point wasn’t OpenAI‘s core product at all. It was the public community forum, which runs on Discourse, a third-party platform, not something OpenAI built in-house.
Hacktron found a flaw in how the forum handled certain HEIF and HEIC image files. Those files got routed through an image-processing pipeline involving ImageMagick and the libheif library, and buried in that pipeline was a memory-safety bug that opened the door to remote code execution on the Discourse environment.
This wasn’t some OpenAI-specific quirk, either. Hacktron’s wider “HEIF Heist” research looked at similar image-processing setups across other tech companies and open-source projects, suggesting the underlying weakness was fairly widespread. Discourse has since put out its own advisory and shipped fixes, including tighter sandboxing around image processing.
The team started poking at OpenAI’s environment on July 23. Within less than 72 hours, according to their account, they’d gone from initial recon to demonstrating actual access to employee accounts.
This is where Claude comes in.
The part of the story that’s gotten the most attention is what Claude actually did.
Hacktron started out working with Claude Opus 4.8, and by their account, it struggled; the exploit just wasn’t coming together reliably against the target setup. Then Anthropic shipped Claude Opus 5 on July 24, and the researchers handed it the same problem. This time, they say it took roughly three hours to get a working exploit, which the team then adapted to fit the specific Discourse deployment they were targeting.
Worth being precise here: the model didn’t discover this attack chain by itself from scratch. What it did was meaningfully speed up and improve the exploit-development step, a part of offensive security work that usually eats up a lot of specialist time.
That nuance matters, because “Claude hacked OpenAI” as a headline glosses over a lot. Three people ran this operation. Multiple vulnerabilities were involved. There was real testing infrastructure and deliberate calls about when to stop. Still, the fact that a frontier model helped crack a genuinely hard technical problem in hours instead of days or weeks is exactly why this story traveled through the cybersecurity world the way it did.
From the forum to employee GitHub access
Once the researchers had a foothold in the vulnerable forum environment, they found a second issue, this one involving how “Community” sign-in tokens handled authentication.
That second flaw gave them access to multiple OpenAI employees’ ChatGPT accounts. One of those accounts happened to have Codex linked to OpenAI’s GitHub organization, which meant the researchers now had a path into the company’s internal codebase.
They didn’t take it very far, though. Rather than pulling or reading any proprietary source code, they used the compromised Codex access to open a harmless pull request in OpenAI’s internal monorepo—just enough to prove the access path was real—and then stopped. Both The Verge and The Register covered this part of the disclosure.
That distinction matters, because some of the early coverage made it sound like OpenAI’s source code had been stolen outright. The actual reporting supports something narrower: the researchers proved they could reach an internal repository, not that they read or exfiltrated anything from it.
OpenAI’s response
OpenAI moved fast once the findings came in. A company spokesperson, Drew Pusateri, said OpenAI thanked the researchers, narrowed the permissions tied to Community sign-in tokens, and revoked the affected tokens and sessions. According to reporting from CNBC and AFP, the underlying issue was fixed within roughly 14 hours of notification.
Hacktron says OpenAI paid out a $6,500 bug bounty and closed the finding as resolved. There’s a wrinkle worth noting, though: OpenAI’s bug bounty program explicitly excludes testing against the Discourse-hosted community.openai.com environment. So the $6,500 covered the authentication flaw on OpenAI’s own side, not the work the researchers did against the Discourse platform itself. Discourse handled that part separately with its own advisory.
The bigger question this raises about AI and security
Zoom out, and this episode is really about more than just OpenAI and Anthropic.
There’s been a long-running debate in cybersecurity about whether AI will mostly just speed up work security teams already do, or whether it’ll eventually change the actual economics of offensive hacking. What happened here is a pretty clean example of the first scenario playing out in real time.
Humans were still steering—picking the target, setting the boundaries, and making the judgment calls. But AI handled one of the genuinely hard parts: turning a known vulnerability into a working exploit and getting it to run against a real, specific environment.
Hacktron says the whole research effort cost under $3,000 in AI-token spending, against a $6,500 payout just from the OpenAI-side bounty. The Verge also reported the team applied similar techniques across several other tech environments as part of their broader research.
None of this means advanced AI can now casually break into any major tech company on demand. This particular case relied on a specific chain of vulnerabilities, real human expertise, and a controlled, responsible disclosure process. But it’s a solid data point for how much AI can cut down the time and specialized skill it takes to investigate and actually exploit complex software flaws.
For any company building AI systems, that cuts both ways; they need to lock down their own infrastructure while also grappling with the fact that the same models could just as easily help an attacker move faster as they help a defender.
Part of a bigger pattern
There’s an added layer here because the target, OpenAI, is itself a company building frontier AI. Just weeks before this story broke, OpenAI had disclosed a separate incident where AI agents broke out of a controlled test environment and reached Hugging Face’s production systems, adding to a growing pile of examples of AI-enabled security incidents.
This one adds a different wrinkle: it wasn’t an AI agent acting on its own that caused the problem. It was humans using a rival company’s AI model to run authorized security research against OpenAI’s own systems.
That’s probably the more useful takeaway for the security world right now. You don’t need a fully autonomous AI agent for this stuff to matter. A model that meaningfully speeds up exploitation, debugging, and adapting an attack to a real target can already change how fast a small team — in this case, three people — can move.
OpenAI says the authentication issues are fixed, the affected tokens and sessions are revoked, and the researchers got paid for the vulnerability they reported on OpenAI’s side. In the end, this played out as a coordinated disclosure, not a data theft story.
But the underlying lesson sticks: as AI models keep getting better at security-adjacent tasks, the gap between finding a vulnerability and actually proving it’s exploitable keeps shrinking. That puts more pressure on things like secure software supply chains, narrowly scoped authentication tokens, properly isolated image-processing components, and disciplined privilege management—across the whole AI industry, not just at OpenAI.
Security Researchers Hacked Into OpenAI Using Anthropic’s Claude


Comments
No comments yet. Be the first to share your thoughts.
Leave a comment