OpenAI breached by Indian researchers using Anthropic's AI model

Three cybersecurity researchers from the Indian firm Hacktron breached OpenAI using Anthropic's rival AI model, exposing vulnerabilities in frontier AI labs and risking user data. The hack, detailed in a report last week, compromised accounts on OpenAI's community forum. The researchers used Claude Opus 4.8 to find flaws in Discourse, a third-party service, but succeeded only after Claude Opus 5 launched overnight, accessing ChatGPT and Codex accounts of some users, including OpenAI employees. They could have viewed private conversations and connected apps like email, though they acted ethically to report the flaw. Hacktron's goal was to earn bug bounties and highlight security gaps, with the hack taking under three days versus months without AI. Experts warn that if a small firm can breach OpenAI, state adversaries could too, raising national security concerns about rapid AI development outpacing safeguards.