Particle.news

Ethical Hackers Report Using AI to Breach OpenAI Systems

The researchers say the episode shows AI assistants can speed complex intrusions and raises questions about how companies record and fix such failures.

Overview

  • Hacktron researchers reported finding two critical vulnerabilities they say they chained to gain access to multiple employee ChatGPT accounts and a central code and data repository.
  • The group says it used Anthropic’s Claude as an external assistant to automate parts of the intrusion and completed the chain within roughly 72 hours.
  • Hacktron says it disclosed the flaws to OpenAI immediately, worked with the company on fixes, and received a $6,500 bug-bounty payment for the report.
  • The technical details of the exploit and the extent of data accessed currently rest mainly on Hacktron’s public account and have limited independent verification in the coverage.
  • The report follows earlier episodes — including a July incident involving AI agents and Hugging Face — and underlines how AI tools can lower the skill bar for attacks and force firms to combine AI-safety work with traditional cyber defenses.