DeepMind Reveals Gemini Robotics 2.5 and Open Safety Benchmark ASIMOV-Jailbreak
DeepMind has announced Gemini Robotics 2.5, an updated foundation model for physical AI systems, alongside the public release of ASIMOV-Jailbreak — an open benchmark designed to evaluate whether robotic AI systems can be manipulated into unsafe behavior. The dual announcement signals a deliberate effort to advance both capability and accountability in embodied AI at the same time, rather than treating them as sequential concerns.
The timing is notable. Physical AI is moving from demonstration environments into operational settings across logistics, manufacturing, and service industries. As deployment scales, the attack surface for adversarial manipulation of robotic systems — through prompt injection, context manipulation, or instruction override — becomes a material risk rather than a theoretical one. DeepMind is treating this as an engineering problem worth instrumentalizing now.
Gemini Robotics 2.5 builds on its predecessor with improved dexterity, task generalization, and instruction-following across varied physical environments. The model is designed to handle multi-step manipulation tasks with less task-specific fine-tuning, reducing the overhead required to deploy capable robotic behavior in new contexts. Underlying the system is tighter integration with DeepMind's broader Gemini model family, allowing the robotics variant to draw on stronger language understanding and spatial reasoning.
ASIMOV-Jailbreak is the more structurally significant release for the industry. The benchmark provides a standardized set of adversarial prompts and manipulation scenarios specifically designed to probe robotic AI for safety failures — cases where a model might be coaxed into ignoring operational constraints, overriding safety boundaries, or executing actions its deployment context prohibits. By open-sourcing the benchmark, DeepMind enables third-party researchers, robotics developers, and enterprise operators to evaluate their own systems against a common standard, rather than relying on proprietary internal testing.
The name references Isaac Asimov's laws of robotics, but the benchmark itself is an empirical tool, not a philosophical one. It operationalizes safety evaluation the same way red-teaming frameworks like those used in large language model evaluation do — by systematically trying to break the system and measuring the results.
For companies integrating robotic systems into operations, the implications are direct. Until now, safety evaluation for physical AI has been fragmented, with each vendor or deployment team developing bespoke testing procedures. ASIMOV-Jailbreak introduces a shared reference point, which matters both for procurement decisions and for regulatory conversations that are beginning to emerge around autonomous physical systems. A common benchmark allows for comparative claims about safety posture, something the industry has lacked.
Gemini Robotics 2.5's improved generalization also affects deployment economics. If the model can be adapted to new task environments with less fine-tuning, the cost and time to operationalize robotic systems in novel settings decreases. This shifts the bottleneck away from model customization and toward physical integration and workflow design — a different kind of problem, but one that organizations are generally better equipped to solve.
The broader signal from this release is that DeepMind is treating physical AI as a systems problem, not just a capability problem. Releasing a safety benchmark alongside a model update suggests an internal view that the credibility of robotic AI deployment depends on the existence of verifiable safety standards — and that waiting for regulators or standards bodies to establish those frameworks is slower than building them directly. Whether the broader ecosystem adopts ASIMOV-Jailbreak as a de facto standard will depend on uptake from other robotics developers, but the infrastructure for that convergence is now publicly available.
Sources: — Ars Technica (https://arstechnica.com/google/2026/11/deepmind-reveals-gemini-robotics-2-5-and-open-safety-benchmark-asimov-jailbreak/)