AI Security Flaw Solved
Is the biggest AI security flaw finally solved? Anthropic's Opus 5 just hit a 0% prompt injection success rate. Here is how they did it.
What the video says
Could the biggest security flaw haunting AI agents finally be solved? Anthropic's newly released Opus-5 model has reportedly hit a 0% prompt injection success rate for browser agents across 129 test scenarios.
The System Card confirms this breakthrough relies on a specific combination of model and software. The 0% rate only holds when auto mode is active.
Without those extra protective layers, the Opus-5 success rate rises to 3.7%. In a surprising twist, the standalone Sonnet 5 actually performs better on its own at 0.93%.
On the broader GraySwan Security benchmark, Opus 5 still leads the industry. It registered a low 2% attacker success rate, beating out both Mythos 5 and Fable 5.
