← November 10, 2026 briefing · added in the 04:00 KST update

Google DeepMind unveils Gemini Robotics 2.5 and robot jailbreak safety benchmark 'ASIMOV-jailbreak' (report)

Announced November 10, 2026

Verified (fact)

According to Ars Technica, Google DeepMind unveiled its robotics model Gemini Robotics 2.5, with success rates reportedly up to twice as high on some tasks. It also released 'ASIMOV-jailbreak,' a public safety benchmark that evaluates how well robot AI withstands jailbreak attacks. Which tasks were measured, what they were compared against, and how the model is offered are not in the collected excerpts.

Needs fact-check (claim)

Up to 2x success rate on some tasks (baseline and conditions not in the excerpts).

Why it matters (interpretation)

When robot models are jailbroken, the harm can be physical. Publishing a jailbreak safety evaluation alongside model improvements could serve as a reference point for physical AI safety evaluation.

Confidence medium · official source pending

Sources

Story thread

  1. 2026-11-10 Google DeepMind unveils Gemini Robotics 2.5 and robot jailbreak safety benchmark 'ASIMOV-jailbreak' (report)
  2. 2026-11-11 Gemini Robotics 2.5 follow-up: humanoid task success rate reaches 96% in a real car factory (report)

More about Google