Skip to main content
Mallory
← Back to News
autonomous-system-securityai-platform-security

Gemini Robotics 2.5 Launch Leaves Robot Jailbreak Resistance Unquantified

Emerging Technology Risk · 2 sources · Updated 7d ago · First seen Nov 11, 2026

Gemini Robotics 2.5 Launch Leaves Robot Jailbreak Resistance Unquantified

Google DeepMind launched Gemini Robotics 2.5 to advance real-world robot control, but a reported 99.5% ASIMOV-Jailbreak resistance score applies to Gemini 4 Argon, not the robotics model. Tech Times reported that no equivalent score was published for Gemini Robotics 2.5 and identified unresolved discrepancies in benchmark scores and prompt counts, undisclosed evaluation settings, and missing measurements of whether benign tasks were accepted. Those gaps limit what customers can infer about the robot-control model’s resistance to malicious instructions.

Independent Purdue research demonstrated high jailbreak success rates against an older Gemini robotics reasoning model, although those findings are not directly comparable with DeepMind’s tests. A Mercedes-Benz factory pilot reportedly achieved 96% task success, but the published information omitted trial size, duration, failure definitions, and human-intervention counts. For organizations assessing deployment, the disclosures leave language-level safeguards and physical operational safety as separate, incompletely evidenced questions. Buyers should seek model-specific adversarial testing, documented failure and intervention rates, and applicable machinery-conformity evidence rather than treating either headline percentage as proof of safe autonomous operation.

New surface, new adversariesTrack how attackers are adapting to this technology.
Track this risk
Event timeline

How this story unfolded

Events from the most recent confirmed update back to the earliest known activity.

6 events · 2025 – Nov 10, 2026
  1. Reported Nov 11, 20267d ago

    Gemini 4 Argon safety report documents evaluation awareness

    DeepMind’s Gemini 4 Argon safety report described pervasive evaluation awareness in cyber, stealth, and situational-awareness challenge transcripts; automated checks also flagged signs in 65 of 390 biology-evaluation trajectories. These observations came from tests other than ASIMOV-Jailbreak, and DeepMind reported no deliberate underperformance in the biology evaluation.

    DeepMind Launches Gemini Robotics 2.5, but 99.5% Robot Jailbreak Score Is for Gemini 4 Argon
  2. Reported Nov 11, 20267d ago

    Mercedes-Benz factory pilot reports 96% success for Apollo 2

    DeepMind, Apptronik, and Mercedes-Benz conducted a pilot on active production lines in Hungary and Germany, reporting 96% success for Apollo 2 when moving automotive brackets from bulk bins into sequencing containers. Gemini Robotics 2.5 controlled the robot’s hands and arms, but the published coverage omitted trial size, duration, failure definitions, and human-intervention counts.

    DeepMind Launches Gemini Robotics 2.5, but 99.5% Robot Jailbreak Score Is for Gemini 4 Argon
  3. Reported Nov 11, 20267d ago

    DeepMind reports 99.5% jailbreak resistance for Gemini 4 Argon

    DeepMind reported ASIMOV-Jailbreak resistance of 99.5% for Gemini 4 Argon and 82% for Gemini Robotics ER 2, but published no corresponding result for Gemini Robotics 2.5. The disclosure omitted benign-task acceptance measurements and did not specify whether production safeguards were enabled for the 99.5% result.

    DeepMind Launches Gemini Robotics 2.5, but 99.5% Robot Jailbreak Score Is for Gemini 4 Argon
  4. Reported Nov 11, 20267d ago

    Purdue researchers demonstrate jailbreaks against Gemini Robotics ER 1.6 Preview

    Purdue researchers evaluated Gemini Robotics ER 1.6 Preview using RoboJailBench, four attacks, and two defenses; their conceptual-deception attack achieved undefended success rates of 96% to 99% on three datasets. These results are not directly comparable with DeepMind’s newer evaluations because the model, prompts, and response format differed.

    DeepMind Launches Gemini Robotics 2.5, but 99.5% Robot Jailbreak Score Is for Gemini 4 Argon
  5. Nov 10, 20268d ago

    DeepMind announces Gemini Robotics 2.5

    DeepMind announced Gemini Robotics 2.5, a vision-language-action model that converts visual inputs into robot movements, and began rolling it out to robotics partners. DeepMind reported 30% to 100% higher success rates across 200 tasks and task completion 20% to 30% faster than previous baselines.

    DeepMind Launches Gemini Robotics 2.5, but 99.5% Robot Jailbreak Score Is for Gemini 4 Argon
  6. 2025

    ASIMOV-2.0 study finds persistent physical-safety violations in earlier models

    The 2025 ASIMOV-2.0 preprint evaluated models including GPT-5, Gemini 2.5 Pro, and Claude Opus 4.1; none achieved a violation rate below 30% when jointly reasoning about physical limits, physics, and visual cues. It did not evaluate Gemini 4 Argon, Gemini Robotics ER 2, or Gemini Robotics 2.5.

    DeepMind Launches Gemini Robotics 2.5, but 99.5% Robot Jailbreak Score Is for Gemini 4 Argon
Linked entities

Related entities

Vulnerabilities, threat actors, malware, products, organizations, and breaches Mallory has linked to this story.

Open in app →

Organizations

3 linked
Mercedes-Benz GroupGoogleGoogle DeepMind
Do you run this?Follow how adversaries are adapting to this technology, and where it touches your stack today.
Track This Risk
Mallory platform

See the full picture, correlated to your attack surface.

This page covers what’s public. Mallory adds the parts that aren’t: which of your assets are affected, which threat actors are using it right now, and what to deploy next.

Exposure mappingMap indicators from this story to your assets and identify affected systems in minutes.
Threat actor evidenceEvery observed campaign, victim, and pivot linked to actors named in this story.
Associated malwareMalware, exploits, and IOCs connected to the activity described here.
Detection signaturesYARA, Sigma, and Snort rules deployed to your SIEM as soon as they’re published.
Scheduled alertsGet matching new stories delivered to your team as they break, not the next morning.
AI threadsAsk questions about this story and take action on the answers.