Today’s theme: capability gains on the software side are colliding with brittle operational systems in the physical world. Meta’s new “frontier” agent model and a supermarket refrigeration hack are the two stories that matter right now — one reshapes who controls personal agents, the other shows the cost when connected control systems fail.

In Brief

Turn a WiFi 7 router into a 320Msps software-defined radio receiver

Why this matters now: The WiFi 7 router SDR hack (w1700k) shows that inexpensive consumer hardware can deliver multi‑hundred‑MHz I/Q captures, opening large‑band spectrum monitoring to hobbyists and small labs.

The maker repo demonstrates how a WiFi 7 router’s 320 MHz channel support and high-speed ADCs can be repurposed to stream wideband I/Q data. That lowers the price barrier for spectrum research: instead of a costly USRP, you can prototype wideband ideas on sub‑$200 hardware. Practical limits are significant, though — vendor filtering, limited dynamic range, firmware constraints and legal/regulatory concerns mean this is a powerful hack, not a turnkey replacement for professional SDR gear. Enthusiasts will likely route around firmware with OpenWrt or tap FPGA interfaces to get usable streams; researchers should still expect to validate performance against purpose-built receivers. Read the maker’s repo at w1700k-sdr on GitHub.

AI coding agents generate more code — but not more shipped software

Why this matters now: The new empirical study summarized by Ars Technica finds that AI agents increase code output but create a bigger review bottleneck, so organizations that celebrate output metrics may be fooling themselves.

Researchers report roughly 30% more lines of code and more commits after agent adoption, but pull requests spend 49% longer in review and the share needing changes nearly doubled. As the researchers put it:

“In particular, code review becomes the bottleneck.”

That matters for engineering managers: raw generation is cheap, but validation, integration and ownership aren’t. The practical fixes are organizational — invest in reviewer capacity, stronger automated testing, and consider AI‑assisted code review tools so generation actually converts to shipped product. See the Ars Technica summary of the study here.

Deep Dive

Muse Wonder

Why this matters now: Meta’s Muse Wonder and Muse Wonder Full are being pitched as “frontier” models tuned for personal and professional agents; they signal Meta’s shift to monetize high‑end agent capabilities and raise immediate privacy and control questions.

Meta quietly introduced Muse Wonder on its research blog as its first models explicitly built for long‑term agent memory, deeper coding tasks and persistent assistant workflows. The feature they emphasize — active memory — lets agents retain and synthesize ongoing context across sessions, which is precisely what makes assistants feel helpful and proactive. Meta claims competitive benchmark performance and highlights agent-centric engineering choices that prioritize memory retention and multi‑turn coherence; the research post frames these as a step toward “true frontier models” for agentic tasks. See Meta’s write‑up at the Meta Research blog.

“We think this is a good default—every Muse user gets a better personal agent as we all collectively use the product and help the model understand the intricacies of human life,” wrote a Meta representative.

That quote crystallizes the tradeoff. Meta’s default setting collects interactions to improve models — a practice many users accept with cloud services but one that raises sharper questions when assistants hold persistent personal memories. The release is also commercial: Muse Wonder will power paid Muse tiers and the Meta Model API, while the beefier Wonder Full appears gated to Muse Code subscribers. That distribution choice matters: instead of an open‑weight release, Meta is signaling it will monetize higher‑end agent capability as a subscription product.

On the trust front, reactions are split. Some in the community applaud another large player pushing agent capabilities forward — better models make agents more useful in productivity workflows — but privacy skeptics point to Meta’s history and ask whether gated weights plus default data collection is the right posture for intimate, always‑on assistants. Policy and product designers should start treating “agent memory” as a feature with its own UX for consent, review, export and deletion: without clear controls, agentization risks concentrating private behavioral data inside a handful of subscription silos. Practically, companies building with Muse Wonder will need to balance capability versus user control, offer transparent opt‑outs, and consider on‑device or client‑side memory primitives where feasible.

Fridges in French supermarkets hit by cyber attack

Why this matters now: A coordinated cyberattack that forced defrosting and heating across supermarket refrigeration systems disrupted food supply and safety, making operational‑technology (OT) security an immediate public‑health and logistics issue.

Overnight incidents reported across France forced major chains — including Intermarché, Coopérative U and Carrefour — to discard tons of fresh and frozen food after remote actors activated defrost or heating modes on display units. Coopérative U’s CEO framed the attack bluntly:

“Targeting equipment essential to the preservation of fresh products and to the food supply of the French people is extremely serious and is unprecedented. It is unacceptable.”

French authorities and ANSSI have opened investigations; retailers took systems offline and coordinated with suppliers to stop further spread. Early reporting names Danfoss AK‑SM series controllers as implicated in some accounts, though attribution and the precise intrusion vector remain under investigation. See reporting at the BBC and a technical summary at Tom's Hardware.

This event highlights a predictable set of OT failures: vendor‑managed IoT exposed to the internet, weak segmentation between retail IT and control networks, legacy controllers with remote management interfaces, and supply‑chain complexity that multiplies blast radius. The consequences weren’t just economic — reports of units reaching extreme temperatures and even fires show how cyber incidents can translate directly into physical danger.

Expect three near‑term effects. First, retailers and suppliers will rush to inventory remote‑management interfaces and apply emergency mitigations: network isolation, default credential changes, patched firmware where available, and temporary removal of remote access. Second, regulators will get involved; public health costs make this a political issue and increase calls for mandatory OT security standards and vendor accountability clauses. Third, insurance and supply‑chain contracts will be revisited: retailers are likely to demand stronger security SLAs from equipment vendors or move to in‑house management where possible. The bottom line: connected control systems are now a mainstream attack surface, and mitigation requires both technical fixes and policy muscle.

Closing Thought

Two truths stood out today: software can get vastly smarter fast — especially around agents — and physical systems stay stubbornly brittle. Where we place control, consent and oversight (in agent memory or in vendor‑managed OT) will determine whether these technologies improve daily life or create new, avoidable harms.

Sources