Seven stories worth your time today, and a theme running through most of them: the bill for everything the industry has been deferring — security, verification, capacity — is starting to arrive, itemized. Walmart is seeing the poor economy catching up to sales, but AI-driven e-commerce sales are up over 23% to keep the company looking good.
If you’re wanting to keep up with AI News that’s updated regularly, check out the newest version of the SkyNet Tribune AI News.
1. Walmart stock drops as sales growth slows
Traffic and ticket sizes came in lower than what Wall Street expected for the quarter. Walmart’s grocery business posted mid-single-digit growth, led by personal care, beauty, and pet supplies, followed by general merchandise, which increased by low single digits.
E-commerce sales were up 23%, above the expected 22% increase, with a 24% increase in the US alone, likely tied to promotions Walmart ran to compete with Amazon’s (AMZN) Prime Day. (Yahoo News)
2. OpenAI is now paying a 20% tax on its own inference
OpenAI disclosed that its expanded monitoring stack adds roughly 20% compute overhead to monitored inference workloads. The safeguards cover all tool-enabled training and evaluation for GPT-5.6 Sol-class models and above, plus all Astra inference: chain-of-thought classifiers watching internal states, escalating to an automated investigator with a 30-minute alert target. OpenAI says it’s absorbing the cost rather than passing it to customers.
The number underneath the number is more interesting. Internal tests on August 7 showed Astra potentially hitting the “critical” tier for cyber risk — capable of finding and exploiting serious flaws in hardened systems unsupervised. OpenAI paused reinforcement-learning training on deployment-bound models for two weeks and put its largest frontier run on hold.
My take: This is the most substantive thing OpenAI has done on safety in a while, and it deserves credit — a two-week pause on a frontier run is real money. But “we’re eating the cost” is a statement about this quarter, not about the equilibrium. A 20% inference tax that scales with capability is a structural cost, and structural costs eventually get passed on or get quietly optimized away. Watch which one happens. (The Register, TNW)
3. Microsoft patched a one-click Copilot data-theft bug — eight months late
Varonis Threat Labs found CoSnitch (CVE-2026-24301), which paired an undocumented autorun=1 URL parameter with the existing q parameter so a single malicious link would silently execute prompts inside a victim’s authenticated Copilot session. The haul: email bodies and metadata, calendar details, Google Drive filenames and summaries, prior conversation history, and saved memory-store instructions.
Varonis reported it on December 31, 2025. Microsoft patched the auto-execution piece on February 1 and shipped the complete fix on August 18, 2026 — roughly seven and a half months. (Nice detail: Varonis found the undocumented parameter by asking Copilot to explain its own security mechanisms, and it volunteered the parameter mid-refusal.)
My take: Eight months is the story, not the CVE. Every vendor shipping an assistant with connectors is building the same shape of vulnerability — an authenticated agent that will do what a URL tells it — and the industry’s disclosure clocks are still calibrated for software that doesn’t have a Gmail token in its pocket. If your org has Copilot connectors live, the honest security question isn’t whether this specific bug is patched. It’s how many autorun-equivalents exist that nobody has looked for. (The Hacker News, Computerworld)
4. Only 3 of 1,357 FDA-authorized AI medical devices were tested on whether patients got better
A University of Toronto team led by Rawan Abulibdeh reviewed every AI-based device the FDA had authorized as of December 5, 2025. Of 1,357 devices, just 34 appeared in registered clinical trials, 12 posted results, 12 produced peer-reviewed manuscripts — and three were evaluated on patient-centered outcomes like mortality or hospitalization. Most trials that did exist excluded pregnant women, adults over 75, and non-English speakers, and ran in well-resourced systems.
My take: This is the clearest number in AI all year, and it should be embarrassing. The authors put it better than I can: clearance tells you a device resembles something already on the market, not that it helps anyone. The 510(k) pathway was designed for devices whose mechanism you could reason about from the predicate. It is a genuinely poor fit for a model whose behavior on your patient population is an empirical question nobody asked. (PLOS Digital Health via News-Medical)
5. Fractile hits a $6.5B valuation on chips that don’t exist yet
The UK inference-silicon startup is in advanced talks to raise ~$600M at a $6.5B pre-money — more than 6x its ~$1B valuation from May. The catalyst: an initial agreement to supply Anthropic roughly $250M in chips. The chips ship in 2027. Founder Walter Goodwin, ex-Oxford robotics; backers include Accel and Founders Fund. Fractile’s pitch is a DRAM-less design that co-locates memory and compute on-die using SRAM instead of shuttling data to separate memory chips.
My take: A 6x re-rate in three months on a pre-revenue contract for unshipped silicon is exactly the kind of thing that looks either visionary or ridiculous depending on a 2027 tape-out. That said, the underlying logic is sound: DRAM is the binding constraint on inference economics right now, and anyone with a credible memory-light architecture has a real buyer. Anthropic isn’t paying $250M for a slide deck. (TNW; architecture background via Tom’s Hardware)
6. Samsung raises foundry prices up to 15%
Per Reuters, Samsung hiked advanced-node pricing on July orders: 10–15% on 4nm (SF4) and 5nm (SF5), ~10% on 8nm. Chinese customers are absorbing the largest increases; Taiwan-based customers saw 5–10%. Demand spilling over from a full TSMC is filling Samsung’s 4nm lines.
My take: When the perpetual number-two in foundry gets pricing power, that’s the tell that capacity is genuinely gone, not just tight. Samsung is deliberately pricing under TSMC’s trajectory — TSMC already moved 5–10% on sub-5nm in January with more signaled for 2027. Everyone downstream of a wafer should be modeling higher COGS through 2027. (Tom’s Hardware)
7. Read Zhipu’s GLM-5.3 cyber numbers past the headline
Z.ai (Zhipu’s international brand) shipped GLM-5.3 this month, claiming a 50% gain over GLM-5.2 on its internal Code Bench plus leading results on Terminal-Bench 3.0. The headline that traveled: 84.5% on CyberGym, edging Anthropic’s Mythos 5 (83.8%) and GPT-5.6 Sol (83.6%). Weights were slated to release two weeks post-launch after additional safety evals.
My take: Scroll one line down and the picture changes. On ExploitBench, GLM-5.3 scores 54.4% against Mythos 5’s 78% and Sol’s 76.5%. That gap is the whole story: GLM-5.3 is competitive at finding vulnerabilities and well behind at autonomously exploiting them. Those are different threat models and they get collapsed into one “beats Western rivals on cyber” headline every single time. The numbers are also self-reported with no disclosed third-party verification. Wait for a red team. (TechNode, Nation Press)
The through-line: OpenAI is paying 20% to watch its own model. Microsoft took eight months to close a one-click exfil path. The FDA has authorized 1,357 devices and verified three. Every one of those is the same story — capability shipped years ahead of the machinery for checking it — and the gap is now expensive enough to show up on a P&L.