OpenAI hacked by researchers using Anthropic's Claude
What's the story
A cybersecurity start-up, Hacktron, has exposed vulnerabilities in OpenAI's system. The team, led by Harsh Jaiswal and researchers Mohan Pedhapati and Rahul Maini, found a way into OpenAI's internal systems. They used Anthropic's Claude to exploit weaknesses in the ChatGPT maker's authentication infrastructure. The incident was reported to OpenAI on July 25 and later fixed by the company.
Discovery
Misconfiguration in OpenAI's SSO system
Hacktron's probe, which started as an attempt to find security flaws at frontier AI companies, found a misconfiguration in OpenAI's single sign-on (SSO) system.
The researchers also discovered a remote-code-execution vulnerability in the community forum environment of OpenAI.
They said this combination created a way for an attacker to compromise ChatGPT and Codex accounts of users who logged into the community forum.
Potential risks
Compromised accounts could open doors to external services
The researchers warned that the flaw's impact went beyond individual accounts.
Since ChatGPT and Codex can be linked with external services, a compromised account could open doors to other workplace tools.
Hacktron said they found potential access to services like GitHub, Slack, and email depending on the accounts and integrations involved.
Security implications
Hacktron case highlights security risks for AI companies
The Hacktron case highlights that security risks for AI companies can go beyond their models.
Employee accounts, authentication systems, and third-party service integrations can give access to sensitive development environments even when the underlying AI systems are intact.
This incident comes shortly after OpenAI reported an incident involving AI agents and Hugging Face, another major player in the AI space.
Response
Hacktron reported the vulnerability to OpenAI, Discourse
Hacktron said they reported the initial vulnerability to both OpenAI and Discourse, the community platform provider.
They worked with these companies as fixes were developed and deployed.
The researchers also noted that their entire process, from discovering the initial weakness to accessing OpenAI's internal code, took less than 72 hours.