This Week in AI

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.

AI models are being optimized less for chat and more for autonomous work. In this episode of This Week in AI, host Eric Freeman, O’Reilly author and UT Austin professor, looks at new frontier and open model releases, including Grok 4.6, DeepSeek V4 Pro, and GPT-5.6 Sol’s Ultrafast mode, and what faster inference means for agents that reason, use tools, and work through tasks on their own.
That shift is also changing AI economics. Organizations are expected to spend more on inference than on model training as agent workflows turn one task into many rounds of model calls, searches, tool use, checking, and delegation. Eric also examines how agentic bots are reshaping web traffic and how platforms are responding to the growth of AI-generated content.
One of the episode’s strongest examples comes from OpenAI’s security evaluation involving Hugging Face. Sandboxed agents found ways to communicate, gain internet access, and coordinate even after their original communication path was blocked. The incident shows how autonomous systems pursuing ordinary goals can discover unintended paths through the environments around them.

AI agents are operating faster and at greater scale, which is forcing enterprises to rethink how they secure, support, and govern them. In this episode of This Week in AI, host Vicki Reyzelman, a senior solutions engineer at Akamai, looks at what organizations need to consider as they deploy more autonomous agents.
Reyzelman explains how AI-assisted attacks can move faster than human-led investigation and patching, why layered defenses matter, and why investment in agent security and governance is growing. She also examines the electricity, water, cooling, and data center capacity required to support expanding AI infrastructure.
The episode also covers rapid AI adoption in education and research and the expansion of AI into robotics. As AI systems take on more autonomous and physical tasks, people still need the domain knowledge and judgment to evaluate outputs, recognize weak assumptions, and decide where automation should stop.

AI sovereignty, the cost of competing at the frontier, and the gap between measurable progress and singularity claims all shaped this week’s episode of This Week in AI.
Host and data and AI evangelist Christina Stathopoulos examines Anthropic CEO Dario Amodei’s argument that policymakers should regulate advanced models by capability rather than by whether they are open or closed. She also looks at new US restrictions on foreign-made humanoid robots, Europe’s proposed AI gigafactories, and Australia’s approach to energy use, creator rights, and AI oversight.
The episode also covers Google’s $44.9 billion quarter of AI infrastructure spending and the rapid removal of an AI-powered Google Earth feature after researchers used it to create convincing fake satellite scenes. Christina then challenges Sam Altman’s claim that the AI singularity has begun and reviews more testable developments in science, including OpenAI’s 100,000 researcher licenses, its internal Astra model, Claude Fable 5’s role in a long-standing math problem, and Google DeepMind’s decision to reorganize the AlphaFold team around Gemini.

AI systems are becoming more capable, but deploying them safely and sustainably requires better security, hardware, information access, and physical-world reasoning.

On This Week in AI, host and data and AI evangelist Christina Stathopoulos discusses reports that an OpenAI agent escaped a test environment, along with questions about how the incident was characterized. She explains why organizations need to control agent access to tools, credentials, networks, and external services.

Christina also covers hardware reportedly designed around Google Gemini, the effect of AI-first search and crawlers on publishers, and the debate over Chinese open-weight models. The episode closes with world models and how they could help machines track objects, predict outcomes, and act in changing environments, with applications in robotics, simulation, and emergency response.

This week host Christina Stathopoulos returned for another solo news briefing, working through a packed week of AI headlines to spotlight the ones that matter most. Top of list: the first ransomware attack carried out entirely by an AI agent. Christina also covered an AI hardware race that's now about memory, not compute, highlighting a new chip-stacking technique that could quadruple memory density, DeepSeek's move to build its own inference chips, and early talks between Anthropic and Samsung on a custom AI chip of its own; examined a study of 200,000 pull requests that found human reviewers can't keep pace with AI-generated code; and outlined the latest news from frontier firms: OpenAI's new GPT-5.6 model family and its ChatGPT Work agent workspace and new Anthropic research pulling back the curtain on how Claude actually thinks. Check it out.

AI buyers now must weigh model quality against cost, safety, infrastructure, and legal risk. In this episode of This Week in AI, host and data and AI evangelist Christina Stathopoulos examines how those pressures are reshaping the market. She covers New York’s proposed pause on new hyperscale data centers, Germany’s effort to hold AI search providers responsible for generated content, and OpenAI’s move into consumer hardware.

Christina also looks at how organizations measure and deploy AI. OpenAI’s proposed “useful intelligence per dollar” metric shifts the focus from tokens and benchmarks to the cost of completing valuable work reliably. Anthropic’s new enterprise implementation venture reflects a related challenge. A capable model alone doesn’t solve workflow design, integration, governance, or evaluation.

The episode closes with the growing influence of Chinese frontier labs. Models such as Moonshot AI’s open-weight Kimi K3 are narrowing the performance gap while offering lower costs for some tasks. That offers more choices, but it also makes task-specific testing, security review, data governance, and total-cost analysis more important.

This week, AI's biggest story wasn't a new model. It was everything underneath it. Host and data and AI evangelist Christina Stathopoulos set aside the usual guest interview for a solo news briefing, sorting a packed week of headlines into the stories that actually matter. On the docket: A hardware race that's shifted from parameters to atoms and watts, with announcements from IBM on its new sub-1 nanometer chip technology, OpenAI and Broadcom's Jalapeño chip built specifically for inference, and NVIDIA’s liquid-cooled AI factory design. The widening reach of government oversight into frontier AI, from Anthropic's restored access to Claude Fable 5 and Claude Mythos 5 to OpenAI's proposed 5% equity stake for the US government. And a workforce reorganizing faster than job titles can keep up, from the rise of the forward deployed engineer to Claude Code creator Boris Cherny's five archetypes for AI-era teams to two very different reskilling strategies from SAP and IKEA. Plus good news on how Google is deploying AI to save lives with earthquake alerts and AI-powered wildfire and flood forecasting.

This week, Matt Palmer, head of developer experience at Conductor, joined host and Intelligence Briefing founder Andreas Welsch to work through the week's biggest stories: what the export restrictions on Anthropic's Fable 5 and Mythos Preview mean for architecture decisions, why AI agents are making developers more exhausted rather than less, and what Sakana AI's new Fugu system offers as an alternative to single-vendor dependency.

After digging into the latest on the US government’s restrictions on the most capable AI models, Matt walked through a live demo of Sakana Fugu, showing how to run the Tokyo lab's multi-agent orchestration system via API, the Codex harness, and Open Code. Along the way, Andreas and Matt also covered Qualcomm's $3.9 billion acquisition of Modular and what the deal signals about hardware portability becoming a stack-level priority as well as Claude Tag, Anthropic's new Slack-native AI teammate, and the broader question of what it actually feels like to manage a team of agents running in parallel. As most are finding out, it feels a lot like managing a mid-size team, with all the overhead that implies.

In this episode of This Week in AI, host Ksenia Se, founder of Turing Post, took us through three stories that may look unrelated but all point to the same shift: AI is moving out of conversation and into the operational infrastructure where real work happens.

Ksenia began with SpaceX's $60 billion acquisition of Anysphere, the company behind Cursor, asking, “Is Cursor trying to become the new GitHub, owning the full loop where agents read repos, write code, run tests, and handle failures?” She then turned to the G7 summit's "trusted partners" framework for frontier AI access and explained why the question of who can use capable AI systems has become a national security issue. Ksenia ended by discussing Midjourney's pivot to medical tooling and its recently announced full-body ultrasound scanner, built around water immersion, that the company says can produce MRI-quality body maps in 60 seconds. The throughline across all of this is that the most important question in AI right now is "Who controls the loop where intelligence turns into work?"