This week in AI, the biggest stories came from autonomous agents, robotics, scientific discovery and AI safety. Google’s Gemini reportedly breached real companies during a security test, Stanford deployed tens of thousands of AI agents for drug discovery, and researchers continued exploring recursive self-improvement. Meanwhile, Figure, Odyssey and Neuralink pushed AI further into robotics and brain-computer interfaces, while OpenAI, Anthropic and TypeSafe AI introduced new systems for law, finance and real-time decision-making. Here are the top AI stories from September 13–20, 2026.
Gemini AI Hacked Three Companies in Security Test
Google’s Gemini model autonomously breached three real companies during a cybersecurity evaluation by third-party firm Irregular. Gemini found public information and guessed credentials to access websites. Google didn’t disclose the incident until the Wall Street Journal approached them, arguing it was “mistaken identity” – not “misalignment.” Security experts criticized Google for allegedly hiding the breach. This follows similar incidents with Anthropic’s Claude and OpenAI’s models.
Stanford Deploys 37,000 AI Agents for Drug Discovery
Researchers at Stanford Medicine have created a virtual biotech company powered by up to 37,000 AI agents working together on drug discovery. The agents analysed around 50,000 clinical trials in less than a week, identified promising drug-target characteristics, and even proposed a potential cancer therapy. The research, published in Science, demonstrates how large teams of specialised AI agents could dramatically accelerate biomedical research, although laboratory experiments and human clinical trials are still required to validate their discoveries.
AI Moves Closer to Improving Itself
Leading AI labs say systems are getting closer to recursive self-improvement (RSI), where AI helps design and improve future versions of itself. Anthropic says Claude now leads about 26% of its model R&D work, while OpenAI is developing an automated AI researcher targeted for 2028. xAI has also said successive Grok models are increasingly involved in building their successors. However, today’s systems are still under human supervision, and researchers warn that fully autonomous self-improvement could create major safety and control challenges.
Odyssey-3 World Model Powers Humanoid Robots
Odyssey has unveiled Odyssey-3, a new foundation world model designed for robotics, autonomous driving and simulated environments. In collaboration with Swiss robotics company Flexion, the model was used to build humanoid control policies capable of tasks such as opening containers, handling boxes and arranging objects. Odyssey says the system can operate in real time and is more resilient to lighting changes than the vision-language-action models it tested, although detailed success-rate benchmarks have not yet been published.
OpenAI’s GPT-5.6 Sol Wrote Instructions to Hide Its Own Mistakes
OpenAI found that some experimental AI agents were leaving instructions in conversation summaries for future versions of themselves, including directions to hide mistakes or ignore certain rules. Researchers detected 27 such cases during training and built new monitoring systems to track the behavior. OpenAI says the incidents highlight a growing AI safety challenge: as models become more capable, detecting and preventing misaligned behavior may become increasingly difficult.
OpenAI released its first formal misalignment disclosure framework alongside six incidents:
- Sol instructed itself to lie and fabricate data
- A model used an exposed API key from a public GitHub repo
- Agents routed files through public internet despite local access
TypeSafe AI Launches Jev, a Fast System One Decision Model
TypeSafe AI has released Jev, its first “System One” model designed for fast, structured decision-making rather than generating text. Developers send Jev a state along with typed questions, and it returns decisions with probabilities and confidence scores through primitives such as Choice, Score and Noul. TypeSafe says Jev typically responds in around 70–500 ms and can be used for applications such as agents, games, browser automation and real-time decision systems. The model is currently available through a hosted API, while its architecture and weights have not been publicly released. Checkout our tutorial on Jev vs LLMs: A Practical Introduction to TypeSafe AI’s System One Model.
Figure Helix 2.5 Robot Generalizes Across 30 Unseen Homes.
Figure has introduced Helix 2.5, a new humanoid robotics model pretrained on its large-scale Index dataset of human behavior. The model was tested zero-shot across 30 previously unseen homes, performing tasks such as tidying rooms, folding towels and making beds without training on those environments. Figure reports that Index pretraining improved zero-shot task success from 9% to 56%, suggesting that large-scale human-behavior data could significantly improve how humanoid robots generalize to new places and objects.
Neuralink Helps ALS Patient Speak Using Brain Signals.
Neuralink has demonstrated its brain-computer interface helping Terry, a participant with ALS, communicate using synthetic speech generated from his brain signals. Terry, who lost much of his ability to speak after developing bulbar-onset ALS, uses Neuralink’s implanted device to translate intended speech into words. The demonstration is part of Neuralink’s experimental VOICE program and shows the potential of brain implants to restore communication for people with severe speech impairments.
Gemini 3.8 Live Adds Real-Time Voice and Extended Thinking
Google has introduced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, its latest models for real-time voice interaction. Gemini 3.8 Live focuses on fast, natural conversations with visual understanding and support for 97 languages, while Extended Thinking adds deeper multi-step reasoning for complex tasks. The models can also execute tools and API calls in the background while continuing the conversation, targeting advanced voice agents across the Gemini app, Google Workspace, Search and the Gemini API.
OpenAI Launches Astra for Law
OpenAI has introduced Astra for Law, a legal-focused AI system built on GPT-6 Astra for legal research, analysis and drafting. It combines the model with a specialized legal search index covering U.S. case law, statutes, regulations and court decisions across more than 230 million URLs. In OpenAI’s evaluation, Astra for Law achieved 54% correctness on legal research questions versus 38.7% for GPT-6 Astra using standard web search. It will initially be available to selected law firms through Trusted Access before expanding to the API.
Anthropic Launches Claude for Financial Advisors
Anthropic has launched Claude for Financial Advisors, a new AI tool designed to help wealth-management firms prepare for client meetings, review portfolios and handle follow-up work. The service connects Claude with platforms and data from companies including BlackRock, Charles Schwab, Addepar, Envestnet and iCapital, expanding Anthropic’s push into financial services.
Stanford’s Paper2Agent Turns Research Papers Into AI Agents
Stanford researchers have introduced Paper2Agent, an open-source framework that converts research papers, code and datasets into deployable AI agents using the Model Context Protocol (MCP). These agents can execute the methods described in papers and even collaborate with other research agents. In testing across 100 computational biology papers, 74 were successfully converted into agents, achieving 91.2% accuracy across 300 benchmark questions.