Anthropic’s Claude helped 3 researchers breach OpenAI in under 72 hours

Anthropic’s Claude helped three security researchers breach OpenAI accounts and reach an internal code repository within 72 hours.

Researchers at cybersecurity startup Hacktron chained an image-processing vulnerability with a flaw in OpenAI’s identity infrastructure in July to gain access to multiple employees’ ChatGPT and Codex accounts.

One compromised Codex account was connected to OpenAI’s GitHub organization, giving the researchers a path into the company’s internal software environment.

The team stopped after instructing the compromised employee’s Codex account to create a harmless pull request inside OpenAI’s private openai/openai monorepo. Hacktron said the researchers did not inspect proprietary source code.

This week, Hacktron disclosed the vulnerabilities and ended further testing.

OpenAI reportedly fixed the identity-side flaw roughly 14 hours after receiving the report and later paid the company a $6,500 bounty.

Anthropic's Opus 5 cleared a hurdle its predecessor could not

The OpenAI attack accelerated after Anthropic released Claude Opus 5, which overcame an exploitation hurdle that its predecessor had repeatedly failed to solve.

Hacktron began examining the image-upload pipeline used by OpenAI’s Discourse community forum on July 23. HEIC and HEIF files were processed through ImageMagick and the underlying libheif decoding library, giving attacker-controlled images a path into vulnerable code.

The researchers supplied Claude Opus 4.8 with a Discourse Docker image and asked it to inspect the installed libheif package for security weaknesses. The model identified missing fixes that left a heap buffer overflow, enabling out-of-bounds reads and writes.

By July 24, Opus 4.8 had produced an exploit that achieved code execution when address space layout randomization (ASLR) was disabled. But repeated attempts to make the exploit work reliably against Discourse’s normal configuration with ASLR enabled failed.

Anthropic released Opus 5 later that day, giving the researchers another route.

Hacktron opened a fresh session with the new model, which produced a working ARM64 exploit for a local Mac within about three hours. The researchers then asked it to adapt the exploit to the x86-64 architecture and jemalloc memory configuration used by Discourse.

By 6 a.m. on July 25, the team had a working exploit that could execute code through a malicious image upload.

With that foothold established, the researchers next tested whether Claude could reproduce the attack against a remote environment with less human intervention.

Hacktron placed the model in an autonomous loop against its own Discourse Cloud instance. The company said Claude initially refused to develop an exploit directly against a remote system, prompting the team to proxy the test environment so it resembled a capture-the-flag security challenge.

Four hours later, the agent had reproduced the attack against the remote test environment.

The researchers then used the resulting exploit against OpenAI’s community forum, where they gained administrative access. A separate weakness in OpenAI’s single-sign-on system allowed them to move from the forum into ChatGPT and Codex accounts.

One compromised employee had connected Codex to OpenAI’s GitHub organization, creating the path the researchers later used to demonstrate access to the company’s internal repository.

Hacktron co-founder Mohan “s1r1us” Pedhapati said the episode showed how quickly AI was compressing exploit-development timelines that once required far more specialized labor.

He said:

“Our main takeaway from hacking OpenAI: AI is reducing the amount of scarce expertise needed to develop exploits. Work that once took months can now take days. Even leading AI labs can be vulnerable.”

However, Hacktron stressed that the operation still depended on experienced human researchers. The company noted:

“This was not completely autonomous hacking, and skilled human guidance remained important.”

Robert Reith, founder of blockchain security firm Accretion, said experienced researchers still supplied much of the judgment needed to turn AI-generated work into a successful attack, but warned that the advantage may erode as models improve.

According to him:

“There’s still a large gap between what skilled researchers + AI can do vs. general population + AI. The scary part is that this gap may become smaller as AI absorbs this knowledge and intuition over time.”

AI Coding agents expand the blast radius of a compromised account

The same coding agents that accelerated the exploit also increased its potential reach once the researchers gained control of an OpenAI employee account.

ChatGPT and Codex can connect to external services, meaning a compromised account may expose whatever integrations a user has authorized. Hacktron cited GitHub, Slack, and email as services that could become reachable, depending on an account’s configuration.

In this case, the employee’s GitHub connection provided the path into OpenAI’s internal repository.

Security agents warned that this concentration of permissions around AI coding tools could make them increasingly attractive targets as Codex, Claude Code and similar agents become more deeply embedded in corporate development workflows.

Codey Blakeney, research lead at Arcee, said:

“The more popular Codex and Claude Code get, the more people are going to try and target them.”

Blakeney said the risk could grow if software development becomes concentrated around a small number of AI providers, creating broader points of failure across engineering teams.

He noted that if regulation moves us to fewer players, it means less choice and more single points of failure. Blakeney added:

“The entire way software engineering works at most places has completely changed with coding agents, and if just one company has a bad day, it’s going to mess up your roadmap and timelines.”

Maxime Fournes, CEO of AI safety advocacy group PauseAI, said the breach also highlighted a longstanding imbalance between attackers and defenders that could become more consequential as AI lowers the cost of developing sophisticated exploits.

According to him, attackers need to find one overlooked weakness, while defenders must secure a much broader attack surface. He noted:

“It’s massively harder and more expensive to defend against all possible flaws than to exploit a single one.”

OpenAI tightened access after the disclosure, while Discourse prepared a patch by July 27 and added further sandboxing around its image-processing system.