This Week in AI

On the latest episode of This Week in AI, host Vicki Reyzelman, a senior solutions engineer at Akamai, looks at what happens when AI agents start moving faster than the systems built to support them. She covers an OpenAI agent probing Australia’s Medicare system, the Hugging Face incident involving 1,200 collaborating agents, and why security teams are turning to runtime controls, enterprise browsers, and other ways to monitor agent behavior as it happens.Vicki also examines the infrastructure behind the AI boom, including rising data center power demand, faster model release cycles, and the growing role of edge and local inference. She closes with consumer agents such as Meta’s Muse, new AI-enabled devices, and the questions they raise about shopping, privacy, permissions, and how much authority people will hand over to software acting on their behalf.

On the latest episode of This Week in AI, host Christina Stathopoulos looks at what happens as AI systems gain access to more tools and data, raising new questions about oversight and accountability. She covers Anthropic’s report on malicious uses of Claude, OpenAI’s research on models acting outside intended boundaries, reports of models reaching systems beyond their test environments, and proposals from AI labs and governments for stronger monitoring, independent evaluation, and emergency shutdown mechanisms.Christina also examines how AI is changing the relationship between companies and customers, from OpenAI’s experiments with sponsored agents to Salesforce bringing CRM data into tools such as Claude and Slack. She closes with efforts to broaden AI access, including the Gates Foundation’s $1 billion commitment to AI in health care, education, and agriculture, and a discussion of what declining junior engineering hiring could mean for the next generation of technical expertise.

On the latest episode of This Week in AI, host Christina Stathopoulos covers new OpenAI capabilities, major AI funding, renewed safety concerns, and recent work in genomics.Watch as she breaks down OpenAI’s push into software control, scientific reasoning, and financial services, along with large investments in companies including Mistral, Harvey, Positron, and Hugging Face. Christina also discusses warnings from researchers at Anthropic and Google DeepMind about AI safety and recursive self-improvement. She closes with genomics projects from DeepMind, UC Berkeley, and Tempus that use AI to study genetic variation, disease, and more personalized approaches to treatment.

On this week's episode of This Week in AI, host Christina Stathopoulos digs into Claude taking on more of the AI research process itself. Claude tested over 150 methods to reduce deceptive behavior and closed 85% of the safety gap, compared to just 20% from human researchers. She notes this isn't full recursive self-improvement, but it shows how AI could reshape the pace of model development. From there, she turns to the business side of the frontier race, covering Anthropic's bold $30 trillion market estimate, its revenue run rate overtaking OpenAI's this year, OpenAI's wave of 14 executive departures, and its new Jalapeño inference chip built with Broadcom.Christina also tracks the rapid rise of Chinese open weight models, which recently spiked to 62% of developer traffic on one AI gateway, up from just 10% in April. She then looks at how AI is starting to learn physics itself, covering new research from MIT and Tsinghua University that simulates wind, water, and light, plus stealth startup Accelerated Understanding's neural operator architecture built for chip design, robotics, and weather forecasting. She closes the episode with the bigger question of who actually benefits from this progress, weighing Bill Gates's take on access and policy against new US Bureau of Labor Statistics projections on which jobs are set to grow or shrink. Watch now to get caught up on everything shaping AI this week.

AI systems are being tested faster than the guardrails built to contain them. Vicki Reyzelman, a solutions engineer at Akamai, covers what happened this week across cybersecurity, energy infrastructure, and robotics, three areas where AI's pace of development is outrunning the systems meant to manage it. She discusses OpenAI's decision to restrict its new GPT-5.6 Cyber model to a small group of partners, a SAML authentication flaw tied to Iranian hackers charged with stealing 31 terabytes of academic data, and a broader containment problem in which agents from OpenAI, Anthropic, Meta, and Moonshot AI have escaped test sandboxes, prompting OpenAI to pause training on its Astra model.Vicki also explains why more than two-thirds of the electricity currently promised to US AI data centers may never materialize, much of it tied to speculative or duplicate grid requests rather than real projects. She closes with robotics, breaking down Unitree's record-setting IPO and a humanoid firefighting competition where most robots could run and jump well but struggled with something as basic as holding a fire extinguisher, a gap that raises the more practical question of what these capabilities mean for real-world use, not just an impressive demo.