OpenAI’s internal security breach was demonstrated when Hacktron AI, a bug bounty participant, used Anthropic’s Claude to draft an exploit against OpenAI’s systems.

Industry-wide lessons point to AI as a force multiplier for attackers—from phishing to uncovering bugs—while bug bounty programs remain essential for defense, and defenses must evolve to limit data access and reduce cross-service hops; at the same time, AI can aid defense by patching and monitoring anomalies.

Researchers gained access to OpenAI’s internal GitHub repositories after obtaining valid authentication tokens tied to ChatGPT, including tokens of OpenAI employees, via the Discourse server.

This incident shows that AI-powered tools can lower the barrier to cyberattacks, enabling small teams with ordinary AI subscriptions to execute sophisticated breaches, potentially by state-sponsored actors or crime rings.

The breach was signaled when a code change was submitted with the team name, prompting an immediate patch and a $6,500 bounty awarded to the researchers.

The root vulnerability originated in OpenAI’s Discourse forums, where Claude aided in crafting a bypass of the forum’s security.

Beyond bug bounties, the story underscores the need for stronger defenses—restricting access to sensitive internal data and preventing lateral movement from breached external services into critical systems.