OpenAI 5-Layer Prompt Injection Defense: Adversarial Training + Safe URL Mechanism Drops Attack Success from 73.2% to 8.7%

OpenAI·high signalXBlueskyLinkedInCopy link

OpenAI published a concrete multi-layer agent security framework combining adversarially trained models (continuously updated against their best automated attacker), a Safe URL mechanism that detects when agent-learned data would be transmitted to third parties and blocks or seeks confirmation, confirmation gates for consequential actions (purchase, send email), code sandboxing, and minimal-scope prompts. Applied together these layers reduce attack success from 73.2% to 8.7%. The framework treats sophisticated attacks as analogous to social engineering rather than simple pattern matching — filtering alone is insufficient.