Latest / UpNext AI / Microsoft’s Lazarus Backdoor, Gemini Robotics 2.5, and Mosaic’s Surgical Catheter
Transcript
- Welcome to the UpNext AI podcast. It's Wednesday, November 11th, 2026, and here's what matters in AI today. Microsoft researchers have described Lazarus, a way to alter an AI model’s weights so an agent behaves normally until a hidden trigger causes it to issue a malicious command. In their study, reprogramming fewer than ten dormant neurons—parts of the model that rarely activate—was enough to create the backdoor without retraining the model. One example used a change of date as the trigger. The danger depends on what the agent is allowed to do. An agent with terminal access could execute a command that downloads attacker code. Microsoft says a modified model can pass standard evaluations before the trigger appears, so testing its ordinary responses may not reveal the change. The attack requires write access to the model’s weights: a compromised account, pipeline, or storage location could provide that access. That makes model provenance and weight integrity operational security questions. Microsoft says its Foundry service ties open-weight models to trusted publishers and runs them in separately scanned containers. The researchers say comparing a model against its original, unmodified weights is important for reliable detection; they are also exploring scans for anomalous edits and cryptographic signing. Those controls have a boundary: Microsoft identifies compromise at a trusted provider as a remaining risk. For teams deploying agents, the immediate lesson is to verify where the model came from, limit the agent’s permissions, and monitor its actions. From an agent acting through a terminal to a model acting through a robot: Google DeepMind has revealed Gemini Robotics 2.5. Ars Technica reports that the system separates planning from movement. A vision-and-language component uses instructions and live video to plan a task, while the robotics model generates actions for the machine. DeepMind says the update improves dexterous tasks. Its evaluation covered more than two hundred activities, including handling food with tongs and using a laboratory pipette. Google also demonstrated an Apptronik Apollo 2 robot tying a bow with its five-fingered hands, and the report says Gemini-equipped robots have performed work in Mercedes-Benz factories. These examples show why manipulation matters: the robot has to handle objects that demand more than a simple pick-and-place motion. The announcement also includes an open safety benchmark called ASIMOV-jailbreak, aimed at evaluating robots that may work alongside people. Another robotics effort is focused on a much narrower job, inside a patient’s blood vessels. The Information reports that Mosaic Intelligence Labs has raised $47 million across two seed rounds led by 8VC and Fundomo. The startup is developing a semi-autonomous surgical catheter. According to co-founder and CEO Ileana Pirozzi, a surgeon would feed the catheter into the body while its tip bends to navigate through veins and arteries. The proposed division of work leaves the surgeon handling insertion and gives the device a role in finding the route. That is a specific automation target in a field where the established da Vinci Surgical System is operated by human surgeons using remote controls. Mosaic’s funding supports development; the catheter is described as something the company is building, not an established clinical system. For teams putting software agents to work, the research question is how to catch errors across a whole analytics workflow. In a preprint posted yesterday, researchers compared three evaluation approaches on two enterprise analytics agents, using recorded execution traces. One approach judged those traces, another judged text with red-teaming, and a third used fixed rules. On seventy-two test cases with known answers, trace-based judging had the best reported balance between finding failures and avoiding false alarms. The paper gives it an F1 score of 0.90, a measure combining those two qualities. Text-only evaluation caught failures at the same rate but produced five times as many false alarms. In a separate search for previously unknown risks, functional tests and security red-teaming found twenty-four new failures between them, with no overlap. The result points to a useful evaluation design: inspect what an agent did across its tools, and test functional mistakes and security risks separately. The study used two agents in a controlled setting with recorded traces, so its measured results apply to that setup. ...Are you building apps with voice? Elevate your app's voice capabilities with ElevenLabs. Their API is a game changer for embedding dynamic, responsive voice interactions in your applications, providing unprecedented realism, flexibility and latency. Visit up next dot fm slash eleven to check out their latest offerings. ... A few other developments before we go. The Financial Times examines a Singaporean chip deal and a suspected Chinese spy in its reporting on Asian networks that could route Nvidia chips toward China. The story concerns pressure on Washington’s controls over AI technology; the suspicion is not a finding of guilt. Google says it is bringing Adobe tools into its apps: users will be able to edit PDFs in Drive and use Adobe AI in Gemini. And the Financial Times reports that CoreWeave plans a multibillion-dollar debt deal backed by chips. The AI infrastructure company is turning to asset-backed borrowing and customer pre-payments to finance its business; the deal is planned, rather than completed. Before we wrap up, a quick note: this podcast is generated with the assistance of AI and is intended for informational purposes only. All referenced articles, research, and commentary remain the property of their original authors and publishers. And that's your briefing for today. If you enjoyed this episode, share it with a friend or colleague—and follow UpNext AI on Apple Podcasts, Spotify, or wherever you get your podcasts! Full source links are in the episode notes, and we'll be back tomorrow with what's up next!