Blog// artificial intelligence

How the Claude AI model breached OpenAI's internal systems in three days

The Hacktron team used Claude Opus 5 to breach OpenAI's forum and internal GitHub in 72 hours for under $3,000 — experts explain why this matters for small businesses too.

19-09-2026 · 10 min read
How the Claude AI model breached OpenAI's internal systems in three days

Security researchers from the Hacktron team used Anthropic's Claude model to breach the internal systems of OpenAI — a company that itself builds competing AI models. As The Decoder reports, the entire process, from initial idea to working access to an internal GitHub code repository, took less than 72 hours. This isn't a story about uniquely gifted hackers: the attack became possible precisely when a new model version, Claude Opus 5, was released.

It might seem irrelevant to a small shop, salon, or manufacturing company in Russia or the CIS that doesn't even have its own development team. But the point isn't whether OpenAI specifically was breached — it's that the cost and speed of sophisticated cyberattacks have dropped sharply. What once required a team of experienced hackers and months of work can now be done by one person with access to a neural network in a couple of days, for a sum that's laughably small by corporate security standards. That changes the risk calculus for any business with a website, a customer account system, an online store, or cloud services.

What actually happened

Three researchers from the Hacktron team gained access to OpenAI employees' working ChatGPT and Codex accounts through the OpenAI community forum (community.openai.com). They then used one of the compromised accounts to get into OpenAI's internal code repository on GitHub.

The attack relied on two linked vulnerabilities. The first was in the libheif library, which processed HEIC-format images uploaded to the forum. According to The Decoder, a fix for this vulnerability had existed in the source code for a year, but no one had flagged it as a security issue. The Debian packages running on the forum server had never received the patch. Using a specially crafted image file, the researchers gained the ability to execute their own code on the server.

The second vulnerability involved OpenAI's central single sign-on (SSO) system. Whoever controlled the forum server could impersonate active forum users and take over their ChatGPT and Codex accounts. The problem extended beyond the forum itself: any other service using sign-in through an OpenAI account would grant the same level of access. To prove the attack was real, the researchers created a harmless pull request in the company's internal repository, posing as an employee — and, they say, did not view any confidential data.

Claude's role: from helpless to a working exploit in three hours

At first the team tried to find and exploit the vulnerability using Claude Opus 4.8. The model built a working exploit, but only with ASLR protection disabled — a standard mechanism that makes memory-based attacks harder. With protection enabled, several sessions failed to produce a reliable version.

On the evening of July 24, Anthropic released Claude Opus 5. According to The Decoder, the new model assembled a working exploit for a local Mac environment in three hours, then adapted it for the server environment running Discourse — the software powering OpenAI's forum. The researchers then ran the model autonomously against their own test setup.

An interesting detail emerged here: the model refused to write an exploit against the real system directly — its built-in safety restrictions kicked in. The researchers then framed the task as a hypothetical performance test, or “benchmark.” Four hours later, the autonomous agent had taken over the test server. On a related task, the team noticed that Claude noticeably outperformed OpenAI's GPT-5.6 Sol.

OpenAI confirmed the fix roughly 14 hours after receiving the report. Discourse, the forum software's developer, responded within a few days.

It's not just about OpenAI

The researchers didn't stop at one breach. The project, dubbed “HEIF Heist,” saw the team test the same attack scheme against Slack, Meta*, GitHub Enterprise, and other major players. The three researchers spent two months on the entire project and less than $3,000 on AI usage. Adapting the attack for each new target took one to two days.

One telling detail: of all the companies tested, only Shopify noticed signs of the attack — despite thousands of uploaded images and repeated file-processing failures at other targets. The other major companies, with serious security teams of their own, simply failed to detect any activity.

The Hacktron team puts it this way: software complexity has long functioned as a kind of natural defense. Even when source code is open and a vulnerability is publicly known, turning it into a working attack tool required rare skills, time, and deep knowledge of the specific system. That complexity was never a real security barrier, but in practice it protected many companies simply because attacks were expensive and labor-intensive. AI removes that protection, replacing scarce expertise with computing power.

Why this matters even without a giant getting hacked

The direct lesson from the OpenAI story isn't “AI is dangerous” — it's that the barrier to entry for sophisticated attacks has dropped sharply. An exploit tailored to a specific outdated library in a specific server configuration used to require a specialist with years of experience. Now a model on the level of Claude Opus 5 does it in hours, and porting the attack to a new target is a day or two of work rather than weeks of hiring a team.

For large companies, this means rethinking threat models and security budgets. For small businesses, it means something else: attacks that used to target only “fat” prizes like OpenAI or Shopify are now becoming economically viable against far more modest sites and services. If hacking an online store with a monthly turnover of a couple million rubles used to simply not be worth it — too expensive and time-consuming for a small haul — the situation is different now. The cost of an attack is falling so much that it pays off even at small scale: stolen customer data, payment details, access to advertising accounts.

Two specific technical facts from the OpenAI story apply to virtually any website:

  • outdated file-processing libraries (for images, documents) are a classic entry point, and a patch may have existed for a year without anyone bothering to apply it;
  • single sign-on (SSO) and account linking between services are convenient for users, but a single breach point grants access to an entire chain of connected services.

What business owners in Russia and the CIS should do this week

Most small companies don't develop complex software of their own and don't maintain an internal code repository like OpenAI's. But they do have a website on a popular CMS, an online store, a customer account system, integrations with CRM, payment systems, and advertising accounts — and the exact same logic applies.

First, check what platform the website runs on and when the CMS, plugins, payment modules, and file-upload modules were last updated. If the site runs on 1C-Bitrix, WordPress, OpenCart, Tilda with custom modules, or a contractor's custom-built solution, ask the developer or hosting provider whether security updates have been applied in the last six months. The libheif story shows that a patch can sit publicly available for a whole year without anyone applying it, simply because “it still works.”

Second, look at where users upload files on the site: photos for reviews, documents for a loan or installment application, files for print production. Exactly these points — processing uploaded images and files — were the entry point for the attack in the OpenAI case. Ask your contractor whether formats and sizes are restricted and whether uploaded files are checked by third-party services, rather than the site simply “accepting everything.”

Third, review which services are linked through single sign-on. If employees log into CRM, email, cloud storage, and advertising accounts using the same Google or Microsoft account, breaching one service opens access to all of them. A simple, free fix here is two-factor authentication wherever it's available, and separating critical access points (payment systems, site admin panel, advertising accounts) from general work accounts.

Fourth, don't rely on the assumption that “we're too small to be noticed.” The Hacktron story shows plainly that of a dozen major companies tested, only Shopify caught signs of the attack. If monitoring systems fail to detect a breach even at giants with dedicated security teams, a small business without such a team has even lower odds of catching a problem on its own. The practical takeaway is not to skimp on basic monitoring: alerts for new admin-panel logins, logs of site changes, and regular backups that are actually tested for restoration, not just stored in an archive.

Fifth, when working with contractors and agencies, ask directly whether they use AI models to test the security of their own solutions, or at least whether they're aware of this practice. Given that the barrier to entry has dropped for defenders as well, it makes sense to ask a website developer for a one-time security audit using modern tools — this is no longer an exclusive service reserved for banks, but an accessible check available for a reasonable price.

Bottom line

The breach of OpenAI by three researchers using Claude isn't a curious anecdote about one AI company embarrassing another. It's a demonstration that the cost of a sophisticated cyberattack has fallen so far that a $3,000 budget and two months of work by three people was enough to test breaches against several tech giants at once — and only one of them noticed. For small businesses, this means that website security, access segregation, and timely updates are no longer a “someday” issue. The economics of attacks have changed, and they now work against any target, not just companies on the scale of OpenAI.

Frequently Asked Questions

What is Claude Opus 5?

Claude Opus 5 is a state-of-the-art AI model developed by Anthropic. It has advanced capabilities in analysis, text generation, and solving complex tasks, making it a powerful tool.

How can AI hack a system?

AI models can analyze large volumes of data, identify vulnerabilities in code and configurations, and automate the exploitation of these vulnerabilities. They can mimic human actions to bypass security mechanisms.

What risks does this pose for small businesses?

For small businesses this means an increased risk of cyberattacks, as AI tools lower the barrier to entry for malicious actors. Vulnerabilities in websites or internal systems can be discovered and exploited faster and more cheaply than before.

  • Meta Platforms Inc. (including social networks Facebook and Instagram*) is recognized as an extremist organization and its activities are prohibited in the Russian Federation.

Based on: The Decoder — Security researchers used Anthropic's Claude to hack OpenAI's internal systems in under 72 hours

Analysis & review by
Alexander K.

Web development, digital advertising and marketing expert since 2012. Founder of AIFORSALE.

How we can help
Online storesMarketingAI agentsContact us
← All articles