prompt injections

OpenAI warns AI browser agents remain vulnerable to prompt injection

The risk has increased with ChatGPT Atlas’ agent mode, which lets the AI view webpages and perform actions like clicking links and typing directly in a user’s browser.