Google DeepMind unveils Gemini Robotics 2.5 with ASIMOV-Jailbreak safety benchmark

Gemini Robotics 2.5 sits at the point where frontier AI models meet physical machines, with Google DeepMind now pairing a robotics release with a named safety benchmark. The appearance of an ASIMOV-Jailbreak benchmark shows that evaluation of robotic systems is becoming part of the product story, not just a research appendix. For readers, this is a sign that robotics is moving closer to mainstream deployment while safety testing is being made more explicit.

Factory and warehouse operators that use robotic systems will care about whether a model can handle routine tasks consistently and safely, because that shapes how much human supervision is still needed. The 96% factory-trial result points to a system being tested in real environments rather than only in demos. If robotics models like Gemini Robotics 2.5 keep improving on dexterity and reliability, builders may be able to automate a wider set of everyday physical jobs.

The harder trade-off is between faster deployment of capable robots and the limits of safety checks in settings where mistakes happen in the physical world. A benchmark named ASIMOV-Jailbreak suggests the industry is also thinking about how robots behave when prompts or instructions are manipulated. That leaves a practical tension for companies: they want broader autonomy, but they also need assurance that systems can be measured against abuse and failure cases.

The total crypto market fell 1.9% in the 4 hours after the first headline, after rising 0.1% in the 2 hours before.

28 prices, one every 30 minutes