1 of 1
Story summary
- OpenAI, the AI research company, acknowledges prompt injection attacks pose a risk to its Atlas AI browser.
- Despite security efforts, OpenAI says these attacks may never be fully mitigated.
- OpenAI has built an automated attacker that uses reinforcement learning to identify vulnerabilities by simulating threats.
- Experts caution Atlas AI browser risks remain high due to access to sensitive data.
- OpenAI recommends limiting access and providing specific mitigation instructions.
