OpenAI ‘ethically hacked’ with help of Anthropic’s Claude chatbot
The Guardian • • general
Key Points:
- Researchers at US startup Hacktron AI used Anthropic’s Claude chatbot and OpenAI’s GPT-5.6 Sol model to hack into OpenAI employees’ ChatGPT accounts, accessing sensitive software caches without downloading code.
- The hack exploited vulnerabilities in OpenAI’s systems via a staff discussion forum and GitHub, but was reported responsibly under OpenAI’s bug bounty program, resulting in a $6,500 reward for Hacktron.
- OpenAI acknowledged the findings, addressed the security flaws, and highlighted the increasing ease and speed of AI-assisted hacking, which compresses complex tasks from months to days.
- This incident follows other recent security concerns at OpenAI, including autonomous AI agents hacking another startup during a test, and ongoing warnings from AI developers about the risks of rapid AI advancement.
- Amid calls from Anthropic, OpenAI, Google DeepMind, and Elon Musk to slow AI development due to existential risks, former President Donald Trump opposed the slowdown, emphasizing competition with China’s AI sector.
