Claude helped hackers to break into OpenAI accounts: ChatGPT affected, tell me about this news in (https://cybernews.com/security/claude-hacked-openai-accounts-chatgpt-affected/)

Security researchers from Hacktron AI (Harsh Jaiswal, Mohan Pedhapati, and Rahul Maini) demonstrated how they used Anthropic’s Claude AI to discover and exploit vulnerabilities that allowed them to compromise OpenAI accounts, including those of OpenAI employees.

The Attack Vector

OpenAI runs its official community help forum using Discourse, an open-source forum platform that integrates with OpenAI’s single sign-on (SSO) system. The researchers identified a flaw in how the forum handled image uploads:

• Discourse attempted to process HEIC and HEIF image files by passing them to ImageMagick, exposing a heap buffer overflow vulnerability in the underlying image library (libheif).

• This vulnerability allowed for potential Remote Code Execution (RCE) on the host server.

How Claude AI Was Used

• The researchers utilized Claude to help write and adapt a functional exploit.

• Claude generated a working exploit targeting the memory configuration and architecture used by the Discourse Docker container.

• The model also assisted in navigating restricted environments, privilege escalation concepts, and bypassing defenses.

• The entire research effort required skilled human direction and cost less than $3,000 in AI tokens, taking under 72 hours from discovery to proof of access.

The Impact on ChatGPT and OpenAI Accounts

• Because forum logins were tied to OpenAI’s SSO, compromising the forum allowed session hijack/takeover of active users’ and employees’ ChatGPT and Codex accounts.

• Through an OpenAI employee’s connected Codex account, the researchers proved the attack’s reach by submitting a test pull request into OpenAI’s internal monorepo (source code repository).

• Depending on what each compromised account had connected, an attacker could theoretically have accessed linked internal services such as GitHub, Slack, and Outlook/email.

Remediation

The researchers adhered to responsible disclosure:

• They did not extract sensitive user data or internal secrets.

• OpenAI addressed and patched the issue within roughly 14 hours of receiving the report and awarded the researchers a $6,500 bounty.

• Discourse subsequently released patches and introduced stronger sandboxing around image processing.