OPENAI SAYS AI BROWSERS MAY ALWAYS BE VULNERABLE TO PROMPT INJECTION ATTACKS

Summary from the TLDR InfoSec newsletter

OpenAI acknowledges that prompt injection attacks in AI browsers like ChatGPT Atlas are a long-term, unsolved security risk, similar to social engineering on the web. The company is layering defenses, including an LLM-based automated attacker that repeatedly probes Atlas in simulation to discover new attack strategies, tighten protections, and enforce user confirmations. Experts still question whether the current agentic browser value justifies their elevated risk.