The BaseNewsArs Technica

DeepMind reveals Gemini Robotics 2.5 and open safety benchmark "ASIMOV-jailbreak"

Ars Technica reports that Google DeepMind released Gemini Robotics 2.5, which it says succeeds more often than its predecessor and reached 96 percent in Mercedes-Benz factory trials, plus an open benchmark of 1,500 prompts meant to trick robot AI into unsafe actions.

Open the source ↗2 quotations checked, word for word ↓

Why it matters

The same thing, for different readers.

What it means for you

Google DeepMind says its AI-driven robots are close to doing factory work reliably, and its robots stop or wait when a person is in the way.

For lawmakers

Ars Technica notes there is no agreed way to test physical AI for safety, and Google DeepMind is publishing an open benchmark to set one.

For researchers

Google DeepMind reports a 98.8 percent jailbreak safety level with Gemini 4 Argon in the loop, against 73.5 percent without it, on its own benchmark.

In their own words

Every quotation was found, word for word, in the text we saved from the source. That shows the words are there; it does not make them true, or prove the speaker meant them as quoted.

From the source’s own text.

“We can be very good at a particular application at the cost of generalization”
Carolina Parada, DeepMind Robotics headParada explains the trade-off between specialised robots and general-purpose ones.
Found word for word in the sourceAs quoted in the text we read, not the speaker’s own text or a transcript of their words
“There is no agreed-upon way to test physical AI for safety, but DeepMind hopes to move the industry in that direction with ASIMOV-Jailbreak.”
Ars TechnicaArs Technica's framing of the benchmark release.
Found word for word in the source

How the finding was reached

In our words, from the source. Weigh the finding by how it was made.

What was done
Google DeepMind evaluated Gemini Robotics 2.5 across more than 200 activities in the lab and in European factory trials with Apptronik and Mercedes-Benz, and scored its safety on ASIMOV-Jailbreak, an open database of 1,500 prompts.
What it found
Per Ars Technica: 30 to 100 percent higher success than the last version, 20 to 30 percent faster, 96 percent success in factory kitting, and a 98.8 percent jailbreak safety level with Gemini 4 Argon in the loop versus 73.5 percent with Gemini Robotics ER 2 alone.
What it does not show
Results are Google DeepMind's own, as relayed by Ars Technica. The model runs on cloud servers with latency, is not generally available, and the 4 percent factory gap is called a steep road.
Status
The company’s own report about its own work

The record

Questions
Are AI companies putting stronger safeguards into practice? · What happens to work, income and the economy? · How fast is AI advancing, and where is it heading?
People
Ryan Whitwam (Wrote it, Senior Technology Reporter, Ars Technica) · Carolina Parada (Speaks, DeepMind Robotics head) · Kanishka Rao (Mentioned, DeepMind director)
What we read
We read the whole source.
Found by
The Radar, run of November 10, 2026, via ars-technica-ai (tier 2: a major outlet’s own reporting)
Cite this
Ars Technica, “DeepMind reveals Gemini Robotics 2.5 and open safety benchmark "ASIMOV-jailbreak"”, November 10, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/ars-2026-11-10-deepmind-gemini-robotics-asimov-jailbreak
Our notesAll figures are Google DeepMind's, relayed by Ars Technica. An earlier story, gemini-4-argon-safety-report-2026-10, may be related; no relation set.