OpenAI Says Prompt Injections a Challenge for AI Browsers, Builds an Attacker to Train ChatGPT Atlas

OpenAI has detailed how prompt injection attacks continue to pose a security challenge for AI browsers like ChatGPT Atlas. To counter this, the company has developed an AI-powered attacker that simulates real-world exploits during training. OpenAI said the approach helps identify weaknesses faster, while acknowledging that prompt injections require continuous mitigation rather than a permanent fix.