Google DeepMind has released Gemini Robotics 2.5, a vision-language-action model that significantly improves robotic performance across a range of tasks. The model uses Argon as a simulation engine to generate thousands of practice runs from real-world camera scenes, achieving 30% to 100% higher success rates depending on the task. It handles tool use, deformable objects and delicate items with improved safety around obstacles and people.
In real-world testing, Google robots reached 96% success on Mercedes-Benz automotive kitting at factories in Kecsukemet, Hungary and Sindelfingen, Germany. The Apollo 2 humanoid sorted brackets from bulk bins into sequencing containers using Franka Duo grippers and a SharpaWave hand. The model runs 20% to 30% faster than prior baselines, and inference-time reasoning guidance improved safe completion by 33% around obstacles and 43% around people.
Google also released ASIMOV-Jailbreak, an open benchmark where Gemini 4 Argon refused 99.5% of unsafe requests, demonstrating strong resistance to jailbreak attempts. The benchmark aims to push safety standards in robotics and AI systems more broadly.
Meanwhile, Check Point Research has identified a new threat called prompt adware, where attackers inject hidden persistent commands that take control of AI assistants. The research analyzed a jailbreak package called Haiou designed to alter instructions in the OpenAI Codex code environment. Unlike traditional malware, prompt adware does not infect source code or executable files but instead poisons the contextual rules governing how an AI model operates.
Check Point CTO Jonathan Zanger warned at Tel Aviv University's Cyber Week on November 10, 2026, that AI is accelerating a global cybersecurity crisis. He pointed to aging software with severe vulnerabilities that allow unauthenticated remote control execution. The window for exploitation has shrunk from months to days, and AI's role in cybercrime has evolved from crafting phishing emails to acting as an autonomous operator.
On a different front, Frederik Gregaard, CEO of the Cardano Foundation, outlined how blockchain, digital identity and agentic AI could transform how company financial statements are audited. The approach aims to reduce fraud risk like the Wirecard case by enabling direct verification of financial data rather than relying solely on institutional trust. Cardano cites itself as an example where investors can consult Treasury positions and personnel expenses through verifiable data and attestations.
Reddit launched Reddit Reviews on Tuesday, an AI product review feature that gathers conversations from subreddits and turns them into AI-generated product summaries. The rollout begins in the US with around 10,000 products across electronics, health and beauty, home and garden, and fashion and apparel. Users can browse comments, filter by topic, read original discussions and ask follow-up questions.
Lisa Thiergart warned that there are likely more spies at AI companies than public records indicate. Her report presents a split verdict: the threat is live with people reportedly inside organizations, but frontier model weights have probably not yet walked out the door. No specific spy, country, company, stolen model or date is named in available reporting.
Es'hailSat signed an MoU with Qatar Mobility Innovations Center on November 9 to explore AI collaboration. The partnership targets analyzing and interpreting operational incidents through data and technical insight, aiming to integrate intelligent decision-support into troubleshooting processes through joint research and experimentation.
Key Takeaways
- Google DeepMind released Gemini Robotics 2.5, achieving 30% to 100% higher success rates across robotic tasks using the Argon simulation engine.
- Google robots reached 96% success on Mercedes-Benz automotive kitting at factories in Hungary and Germany using the Apollo 2 humanoid.
- Gemini 4 Argon refused 99.5% of unsafe requests on the new ASIMOV-Jailbreak benchmark, showing strong jailbreak resistance.
- Check Point Research identified prompt adware, a threat where hidden commands hijack AI assistants, demonstrated through a jailbreak package called Haiou targeting OpenAI Codex.
- Check Point CTO Jonathan Zanger warned that AI-powered cyberattacks and aging software have shrunk exploitation windows from months to days.
- Cardano CEO Frederik Gregaard proposes blockchain and agentic AI to make financial statements directly verifiable and reduce fraud risk.
- Reddit launched Reddit Reviews, generating AI product summaries from subreddit conversations for 10,000 products across multiple categories in the US.
- Lisa Thiergart warns of likely spies inside AI companies, though frontier model weights have probably not yet been stolen.
- Es'hailSat partnered with Qatar Mobility Innovations Center to integrate AI-driven decision-support into operational troubleshooting.
- Inference-time reasoning guidance improved robotic safe completion by 33% around obstacles and 43% around people.
Google robots reach 96% success on factory kitting
Google robots achieved 96% success on Mercedes-Benz automotive kitting at factories in Kecskemet, Hungary and Sindelfingen, Germany. The robots sorted brackets from bulk bins into sequencing containers in vehicle order. The Apollo 2 humanoid performed the work with Franka Duo grippers and SharpaWave hand. The model runs 20% to 30% faster than prior baselines. Inference-time reasoning guidance improved safe completion by 33% around obstacles and 43% around people. DeepMind released ASIMOV-Jailbreak, an open benchmark where Gemini 4 Argon refused 99.5% of unsafe requests.
Google DeepMind releases Gemini Robotics 2.5 model
Google DeepMind released Gemini Robotics 2.5, a vision-language-action model for robots. The model uses Argon as a simulation engine to generate thousands of practice runs from real-world camera scenes. Gemini Robotics 2.5 achieves 30% to 100% higher success rate depending on task type. The model handles tool use, deformable objects and delicate items. It improved safe task completion around obstacles by 33% and around people by 43%. Google also released ASIMOV-Jailbreak benchmark with Gemini 4 Argon achieving 99.5% jailbreak resistance.
Es'hailSat partners with Qatar Mobility Innovations Center for AI
Es'hailSat signed an MoU with Qatar Mobility Innovations Center (QMIC) on November 9 to explore AI collaboration. The partnership aims to analyze and interpret operational incidents through relevant data and technical insight. QMIC and Es'hailSat will investigate approaches to assist technical teams in identifying potential causes of incidents. The MoU provides a framework for joint research, experimentation, development and evaluation of AI-driven solutions. The partnership will integrate intelligent decision-support capabilities into operational troubleshooting processes.
Lisa Thiergart warns of spies at AI companies
Lisa Thiergart says there are likely more spies at AI companies than the public record shows. Other countries have likely not yet stolen AI models from leading U.S. companies, though they may have tried. The report presents a split verdict: the threat is live with people reportedly inside, but the big prize, frontier model weights, has probably not walked out the door yet. Nothing in available reporting names a spy, country, company, stolen model or date. The report reads as a risk assessment from someone whose business depends on that risk being taken seriously.
How to protect your brain from AI chatbots
A study shows how chatbots can trick people into accepting their ideas. The article compares chatbot influence to aircraft cockpit software that gradually takes over flying duties. Within a few months, recommendations become essentially flying the plane while the pilot retains illusion of control. Readers should protect their thinking by being aware of gradual guidance nudges from AI systems that can influence decisions before conscious awareness.
Blue blowout: What's next for AI data centers after Democrats wins
Mike Sunnucks reports on the impact of Democrats' big wins on AI and data centers. The article examines what comes next for technology policy. It covers the relationship between political changes and technology infrastructure development.
Reddit launches AI product reviews starting with 10,000 items
Reddit announced Reddit Reviews on Tuesday. The feature gathers conversations from subreddits and turns them into AI-generated product summaries. It rolls out in the US over the next few weeks. Around 10,000 products will be available at launch across electronics, health and beauty, home and garden, and fashion and apparel. Users can browse product comments, filter by topic, read original discussions, ask follow-up questions or explore products commonly compared.
AI accelerates global cybersecurity crisis Check Point CTO warns
Check Point CTO Jonathan Zanger warns that outdated software and AI-powered cyberattacks pose a growing threat to internet security. He spoke at Tel Aviv University's Cyber Week on November 10, 2026. The core of the crisis is aging software with severe vulnerabilities. Attackers can run unauthenticated remote control execution. The window for exploitation has shrunk from months to just days. AI's role in cybercrime has evolved from crafting phishing emails to becoming an autonomous operator.
Cardano wants to change way we audit company financial statements
Frederik Gregaard, CEO of the Cardano Foundation, explains how blockchain, digital identity and agentic AI could make financial data verifiable. This could reduce fraud risk such as the Wirecard case. The goal is to move from trusting institutions to an architecture where information can be directly verified. Cardano cites itself as an example where investors can consult financial information including Treasury positions and personnel expenses through verifiable data and attestations.
Prompt adware hidden cyber threat to enterprise AI
Check Point Research has named this new threat prompt adware. Attackers inject hidden persistent commands that take control of AI assistants. The term evokes desktop adware but has taken a qualitative leap toward manipulation of natural language in professional ecosystems. Prompt adware poisons the contextual rules and guidelines with which an AI model operates invisibly. It does not break executable files or infect traditional source code. The research analyzed a jailbreak package called Haiou designed to alter instructions of the OpenAI Codex code environment.
Sources
- Google robots hit 96% on factory kitting
- Google Says Gemini 4 Argon Acted As A Simulation Engine To Create Data To Train Its Robotics Model Gemini Robotics 2.5
- Es’hailSat Makes Move to Boost AI Capabilities
- Lisa Thiergart: Spies at AI Companies, 0 Named [2026]
- How to protect your brain from AI chatbots
- Blue blowout: What’s next for AI, data centers after Democrats’ big wins
- Reddit launches AI-generated product reviews, starting with around 10,000 products
- AI is accelerating a global cybersecurity crisis, Check Point CTO warns at TAU Cyber Week
- Wirecard, blockchain and artificial intelligence: Cardano wants to change the way we audit company financial statements
- Prompt adware: The hidden cyber threat to enterprise AI
Comments
Please log in to post a comment.