Continuously hardening ChatGPT Atlas against prompt injection

In three linesOpenAI hardens ChatGPT Atlas against prompt injection attacks using automated red teaming with reinforcement learning. A continuous discover-and-patch loop identifies novel exploits early and strengthens the browser agent's defenses.

Summary generated by Claude — human-verified