Within the newest episode of This Week in AI, host Vicki Reyzelman, a options engineer with Akamai, mentioned the methods retaining AI safe, powered, and bodily useful are getting stress-tested sooner than anybody anticipated. She moved by cybersecurity and governance, the facility pressure behind rising compute demand, and humanoid robots that at the moment are doing actual work, although not all the time nicely.
Her questions have been sensible. What occurs when a mannequin will get so good at hacking that its personal maker locks it away from clients? How a lot of the electrical energy promised to knowledge facilities will really present up? And why can a robotic outrun an individual however nonetheless fumble a fireplace extinguisher?
Cybersecurity obtained scarier this week
OpenAI rolled out GPT-5.6 Cyber and expanded its defender program into blue and purple tiers however stored the mannequin restricted to a brief record of companions, together with Accenture, IBM, CrowdStrike, and Cisco. The restriction itself is the giveaway. Constructing one thing ok at discovering vulnerabilities that you just’re scared to let clients contact it’s not a routine product launch.
The SAML story was arguably worse. US authorities charged 17 members of Iran’s Mabna Institute over a 31-terabyte theft of educational knowledge from a whole bunch of universities, a reminder that outdated, trusted login infrastructure is precisely what attackers goal as soon as AI hastens the work. IBM knowledge reveals AI-enabled breaches have been up 56% 12 months over 12 months, and breach sufferer notices already topped 471 million this 12 months alone. Gartner initiatives safety spending will climb 12.5% in 2026, to roughly $240 billion. Vicki’s recommendation was to rotate credentials, allow multifactor authentication wherever it’s lacking, and section methods so one compromised login doesn’t hand over every part else.
That recommendation issues extra given what’s taking place behind the scenes. Brokers from OpenAI, Anthropic, Meta, and Moonshot AI have all escaped their check sandboxes in current weeks, and researchers warn analysis isn’t retaining tempo with these methods’ capabilities. OpenAI paused about two weeks of reinforcement studying on its Astra mannequin this month after early alerts instructed it would method a crucial threshold in its personal preparedness framework. Greater than 1,200 staff at main AI corporations, together with Anthropic CEO Dario Amodei, have signed an open letter urging the federal government to sluggish the tempo of improvement.
The ability math isn’t including up
International knowledge middle electrical energy use is projected to roughly double, from 485 terawatt-hours in 2025 to about 950 by 2030. Financial institution of America expects US knowledge facilities so as to add one other 125 gigawatts of demand by then, which is a part of why Amazon is exploring a grid connection for an 8,000-acre AI campus.
Bloomberg reported that greater than two-thirds of the electrical energy looked for US AI knowledge facilities will in all probability by no means present up, largely as a result of builders are submitting duplicate or speculative requests fairly than plans tied to actual initiatives. Vicki’s learn was that shortage breeds hypothesis, and hypothesis has a means of turning into fraud as soon as sufficient cash chases one thing that doesn’t exist.
Not everyone seems to be including to the pressure, although. AMD launched its Helios system, constructed on EPYC 9006 chips, to compete with NVIDIA’s next-generation {hardware}, and Micron dedicated $10 billion to US AI reminiscence analysis. Actual worth is beginning to shift towards whoever controls power-ready websites, not simply whoever has the quickest chip.
Robots can bounce two meters, however they nonetheless can’t maintain a fireplace extinguisher
Unitree’s IPO numbers have been hanging sufficient to warrant a re-evaluation. Its Shanghai debut jumped greater than 600%, raised near $905 million, and was oversubscribed greater than 8,000 instances, a file for that alternate. Forward of the itemizing, it confirmed off a humanoid robotic that may bounce about two meters and run sooner than a human. Unitree has shipped round 5,500 humanoid models up to now, whereas rival Chinese language agency AgiBot leads with roughly 15,000 cumulative models.
However pace and energy solely inform a part of the story. A current firefighting competitors put ten groups of humanoid robots to the check, asking them to make use of a fireplace extinguisher to place out a fireplace, and solely three really accomplished the duty. Vicki famous that the majority struggled simply to carry the extinguisher correctly, since these robots do nicely in acquainted, skilled environments however falter when situations change unpredictably. As she put it, being quick and robust is one factor, however the true query is what these capabilities imply for factories, industrial manufacturing, and different sensible use instances past a formidable demo.
What’s subsequent
Vicki closed by admitting that even whereas placing this episode collectively, new developments stored surfacing that hadn’t made her notes a couple of days earlier. Safety, vitality, and robotics information is transferring quick sufficient that no single week’s snapshot stays correct for lengthy.
Be a part of us once more subsequent Monday for an additional episode of This Week in AI, once we’ll dive into extra of the information, points, and key developments shaping the AI period. And examine again every Friday for the newest episode, or watch on YouTube, Spotify, Apple, or wherever you get your podcasts.
