Researchers used Claude to breach OpenAI's private codebase

A small team of researchers revealed that they identified vulnerabilities in OpenAI's system. The researchers published a blog, where they revealed that OpenAI paid them a $6,500 bug bounty to highlight the weaknesses.

A security investigation into leading artificial intelligence companies has uncovered a route into OpenAI’s internal systems, with researchers using Anthropic’s Claude to help exploit weaknesses in the ChatGPT maker’s authentication infrastructure.

The work was carried out by Hacktron, a small cybersecurity start-up led by Harsh Jaiswal with researchers Mohan Pedhapati and Rahul Maini. The team said it was able to compromise multiple OpenAI employee ChatGPT accounts and use the resulting access to interact with software hosted in the company’s private GitHub environment.

The incident took place on July 25 and was disclosed to OpenAI rather than used for further access. The company subsequently fixed the vulnerabilities and paid the researchers a $6,500 bounty, according to Hacktron.

An OpenAI spokesperson said the problem had been resolved and thanked the researchers for bringing it to the company’s attention, according to The Information.

Weakness in identity system opened wider access

Hacktron said its investigation began as an effort to identify security weaknesses at companies developing frontier AI systems. Researchers eventually uncovered a misconfiguration involving OpenAI’s single sign-on, or SSO, infrastructure, alongside a remote-code-execution vulnerability in OpenAI’s community forum environment.

According to the researchers, the combination created a route that could allow an attacker to take control of ChatGPT and Codex accounts belonging to people who signed into OpenAI’s community forum.

The significance of the flaw extended beyond those accounts themselves. ChatGPT and Codex can be connected to external services, meaning a compromised account could potentially provide a pathway towards other workplace tools. Hacktron said the researchers identified possible access to services including GitHub, Slack and email, depending on the accounts and integrations involved.

In their testing, the researchers reached OpenAI’s private software environment and were able to read code and propose changes. They said the entire process, from discovering the initial weakness to reaching the company’s internal code, took less than 72 hours.

Hacktron said it reported the initial vulnerability to both OpenAI and Discourse, the company behind the community platform, and worked with them while fixes were developed and deployed.

“We immediately reported the initial vulnerability to OpenAI and Discourse and worked with them to coordinate the patch,” the researchers said in their disclosure. They added that OpenAI paid them a $6,500 bounty.

Growing concern around AI-assisted attacks

The disclosure comes shortly after OpenAI revealed an incident involving AI agents and Hugging Face , another prominent organisation in the AI ecosystem. That episode highlighted concerns increasingly capable AI systems being used to conduct cyber operations.

The Hacktron case presents a different scenario. Rather than compromising an AI model itself, the researchers used Claude as part of an operation targeting weaknesses around OpenAI’s services and identity controls.

It also illustrates why security risks for AI companies can extend beyond their models. Employee accounts, authentication systems and integrations with third-party services can provide access to sensitive development environments even when the underlying AI systems remain intact.

Hacktron said OpenAI and Discourse responded to the reported weaknesses and that the relevant issues had been addressed. The researchers’ disclosure therefore describes a vulnerability that was identified and patched rather than an ongoing breach.

researchers used claude to breach openais private codebase

Palantir CEO Alex Karp says AI labs may seek government protection from copyright lawsuits

Australia considers banning the use of smart glasses in government buildings

OpenAI, Microsoft copyright lawsuit: NYT filing reveals paywall bypass claims

Altman, Cook and Huang expected to join Trump, Xi at state dinner

California Governor Gavin Newsom mulls special session or executive action as AI safety concerns grow: Report