Researchers Reportedly Use Claude Opus 5 to Chain Flaws and Access OpenAI Employee Accounts

According to a report published by The Hacker News on September 19, 2026, three researchers at the security firm Hacktron allege they used Anthropic's Claude Opus 5 large language model to chain two flaws and gain unauthorized access to OpenAI employee accounts and internal systems. The chain is said to have started with a bug in the software running OpenAI's public help forum, which was then moved through a weakness in OpenAI's own login system. The researchers claim the access extended to ChatGPT and Codex accounts belonging to several OpenAI employees, as well as an internal OpenAI code repository.

The report characterizes the activity as security research. However, the publication does not provide detailed technical descriptions of the alleged flaws, nor does it publish a proof-of-concept or disclose the specific software versions involved. No CVE identifiers have been assigned, and the report notes that no patches were specified, though it suggests OpenAI likely addressed the issues after disclosure. Anthropic and OpenAI have not issued official statements confirming or denying the claims. The veracity of the assertion that Claude Opus 5 was the enabling tool for the chaining process remains unconfirmed, and the story has not been independently verified at the time of writing.