Type: url
Original: https://openai.com/index/designing-agents-to-resist-prompt-injection/
Status: generated
Summary
An 11 March 2026 OpenAI post arguing that the most effective real-world prompt injection attacks have come to resemble social engineering rather than simple instruction override, and that defence should therefore constrain the impact of manipulation rather than try to classify malicious input. It describes source-sink analysis as the complementary framing and Safe Url as the deployed mitigation against conversational data being transmitted to third parties.