OpenAI is using automated red teaming to fight prompt injections in ChatGPT Atlas. The company compares the problem to online fraud against humans, a framing that downplays a technical flaw that could slow the rise of the agentic web.