Breaking

Researchers Used Claude to Reach OpenAI Internal Code in Bug-Bounty Test

September 17, 2026 at 22:41 EDT

Security
AI Agents
Software Dev & Coding

Independent security researchers used Anthropic's Claude to gain access to an OpenAI employee's ChatGPT account and reach private internal code, exposing how AI agents can accelerate both legitimate bug hunting and dangerous intrusions. OpenAI confirmed the issues were resolved and paid the researchers a $6,500 bounty after they reported the vulnerability through its bug-bounty process.

Sep. 18, 2026 · AI security

Claude-assisted test reached OpenAI code

Independent researchers chained a forum flaw and authentication-token access into a private-code reach during a bug-bounty test. The case shows both the defensive value and the containment risk of AI agents in security work.

$6,500

Bounty paid after disclosure

Bug bounties in BountyBench

90%

Top patch score in benchmark

Patch performance, drawn to scale

90%

Top agents

87.5%

Claude Code

57.5%

Claude Code exploit

The same class of agents can help repair systems and probe them, making containment and credential scoping central controls.

How the access chain unfolded

Forum flaw

→

Tokens obtained

→

Cross-service access

→

Harmless proof

Defensive upside

AI-assisted testing can find serious identity and access-management failures faster when paired with responsible disclosure.

Operational risk

The same workflow can lower the skill needed to chain vulnerabilities if credentials are overbroad or agents reach live systems.

Continue reading

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