University of Washington study reveals prompt injection risks lurking in AI agent memory
A University of Washington study published on July 16, 2026, as arXiv:2607.14611 found that AI agents from Anthropic and OpenAI can correctly refuse malicious instructions but still store them in pers…