TechSecurity researchers used Claude to help hack into OpenAI

Three independent Hacktron researchers say they compromised OpenAI employee accounts in under 72 hours with Anthropic’s Claude Opus 4.8 and 5. They reached OpenAI’s GitHub “Monorepo” through a HEIF-processing flaw in Discourse, but did not access its internal code. A Codex pull request demonstrated the breach. The vulnerabilities were fixed, and OpenAI paid Hacktron $6,500.

OpenAIcybersecurityClaudevulnerability

TThe Verge★★★☆☆2026-09-18 15:30Original