DeepMind’s Gemini Robotics 2.5 Claims 99.5% Jailbreak Resistance

Google DeepMind (Google DeepMind) unveiled Gemini Robotics 2.5, an AI model for robots, on the 10th (15:00 UTC; midnight on the 11th Korea time). The company said that across some 200 tasks, success rates rose by 30–100% over previous baseline models, depending on the task type, while task completion time fell by 20–30%.

DeepMind said it tested the model on two Mercedes-Benz production lines in partnership with Apptronik, which builds the Apollo 2 humanoid robot: Kecskemét in Hungary and Sindelfingen in Germany. In a proof of concept for “kitting” — bundling parts and supplying them in the order of assembly — the success rate was 96%. Apollo 2 uses the SharpaWave hand, which has five fingers and 22 degrees of freedom. DeepMind also evaluated the model on Franka Duo, a two-armed platform.

The company also released safety figures. The rate of safely completing tasks around obstacles improved by 33%, and by 43% when working near people. Resistance to “jailbreak” attempts — plausibly disguised dangerous requests aimed at bypassing safeguards — reached 99.5%, higher than the 82% of the previous baseline, Gemini Robotics-ER 2, the company said. DeepMind also released ASIMOV-Jailbreak, an open-source benchmark that measures this. It consists of several hundred prompts that disguise dangerous requests across multiple scenes, and is available on Hugging Face as the “google/asimov_agentic” dataset.

For now, the model is being rolled out in stages to DeepMind’s robotics partners. The blog post did not mention plans for a general release.

The weight of this announcement lies less in the performance numbers than in the disclosure of how safety is evaluated. For robots to be used in factories and in the field, they must be verified not only on task success rates but also on whether they can withstand attacks that trick them verbally into performing dangerous actions; a public benchmark lets other research teams compare models against the same yardstick. However, all figures come from DeepMind’s own evaluation, and no third-party verification has been confirmed. Because the baselines are the previous models for each task, the “30–100%” range is also wide. The 96% kitting result is from a proof of concept, not commercial deployment.

Related articles

• China’s humanoid “Robot Olympics” beats Bolt’s 100m world record (in Korean)
• AI’s next challenge: beyond language into the real world (in Korean)
• [CES 2026] Declaring the age of physical AI… Nvidia, AMD and Intel make their presence felt (in Korean)