Opus 5: AI Prompt Injection Solved by Anthropic?

Opus 5 may have solved browser-based prompt injection, a major security flaw for AI agents. Anthropic states Opus 5 is nearly immune to these injections within its own software. For browser agents, the attack success rate was zero percent across 129 test scenarios. This is significant because prompt injection can allow attackers to bypass an AI model's instructions using manipulated inputs. Here's the thing: this zero percent rate is achieved when "Auto Mode" is active. Auto Mode uses two defense layers. One layer scans incoming data for hidden instructions. The other blocks dangerous actions before they can be executed. An attacker would need to defeat both layers independently. Without Auto Mode, Opus 5's success rate against attacks is 3.7 percent. This shows that the combination of the model and its protective software is key to achieving this high level of security. This could mean a more secure future for AI interactions online.