Breaking

Researchers Used Anthropic's Claude to Access an OpenAI Employee's ChatGPT Account

September 18, 2026 at 00:23 EDT

Security
AI Agents
Software Dev & Coding

Security researchers at Hacktron AI used Anthropic's Claude to exploit weaknesses around OpenAI's Discourse-hosted community forum, then used resulting sign-in tokens to reach an OpenAI employee's ChatGPT account and internal GitHub access. OpenAI fixed the issues, revoked affected sessions and paid a $6,500 bug bounty, but the episode turns AI-assisted coding progress into a concrete identity-security test.

Sep. 18, 2026 · OpenAI security

Claude-assisted account breach

A three-person research team used Anthropic's model to turn a third-party flaw into employee-account and code-repository access. The case was resolved through a bug bounty, but it exposed a wider identity-security problem.

$6,500

bug bounty paid

researchers involved

26%

Claude-led R&D share

90%

Claude-collaborative R&D

How much AI is in the research loop?

26%

led by Claude

90%

done with Claude

The larger share is collaboration under human direction; the smaller share is work Claude can mostly carry from a high-level prompt.

The access chain

Image-file flaw

→

Tokens exposed

→

ChatGPT account

→

GitHub access

Defensive upside

Agentic coding models can speed up legitimate bug discovery, root-cause analysis and proof-of-concept testing when researchers stay inside disclosure rules.

Operational risk

A weak token boundary can turn a narrow service bug into access across higher-value tools, leaving defenders with less time to react.

The lesson is not just stronger models

It is tighter scoping: forum tokens, ChatGPT sessions and code-repository permissions cannot be treated as separate risks when agents can chain them together.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.