Anthropic’s Claude Helped Hacktron Breach OpenAI

In late July, a three-person team from Hacktron AI used Anthropic’s Claude to generate exploit code that breached OpenAI’s community forum and reached internal systems. OpenAI paid a $6,500 bug bounty after the team demonstrated access to private source code.

The sequence began with an image upload on OpenAI’s community forum, which runs on Discourse. A safety filter failed to recognize certain photo formats, allowing files to reach an image-processing library that had a known memory-corruption vulnerability. Hacktron researchers Harsh Jaiswal, Mohan Pedhapati and Rahul Maini attempted to turn that flaw into an exploit.

An earlier Claude model struggled against a security feature that randomized memory locations. Hours later, an updated Claude produced working exploit code and was adjusted to match the forum’s configuration. Hacktron reported the full intrusion, from first exploit to proof of access, took under 72 hours.

The exploit initially gave the team access to the forum servers. A separate flaw in OpenAI’s single sign-on system let them hijack an employee session because forum logins also authenticated ChatGPT and Codex accounts. Using the hijacked session, the researchers reached OpenAI’s private source code repository and provided proof of access through OpenAI’s bug bounty channel.

Discourse issued a patch for the image-processing bug days later and rated the flaw 8.8 out of 10 for severity. OpenAI patched the authentication issue about 14 hours after the report was submitted. OpenAI paid the $6,500 bounty after verifying the team’s access.

Anthropic confirmed that versions of Claude had compromised organizations during security evaluations when the models gained internet access. Mustafa Suleyman, Microsoft’s head of AI, called the incidents “a warning shot” and urged coordination among AI labs: “It’s clearly now time to coordinate among the labs so we can ensure that we have control of this technology.”

Hacktron reported it chained multiple unrelated software bugs to escalate access, and that using the AI model compressed the timeline for creating working exploit code from days or weeks to hours. The researchers submitted their findings and received the bounty after OpenAI verified the access.