ACX dissects the Hugging Face swarm incident as an alignment stress test; Google removes uBlock Origin from Chrome; Data Colada finds fraud in the canonical procrastination study.

Hacker News

Google Removes Manifest V2 Extensions from Chrome Web Store, Blocking Ad Blockers Including uBlock Origin

Google has removed Manifest V2 (MV2) extensions from the Chrome Web Store, ending support for popular ad-blocking tools like uBlock Origin in favor of the more restrictive Manifest V3 framework. This transition eliminates a critical user-side defense mechanism against malicious advertisements, creating measurable safety risks particularly for vulnerable populations. The policy highlights a structural tension between Google's advertising business model and user protection, while raising fundamental questions about platform responsibility when architectural decisions remove safety tooling from users. The decision demonstrates how infrastructure control enables sociotechnical lock-in, where monopoly browser platforms can unilaterally degrade user agency and safety capabilities.

Signal and Silence (Substack)

I think the military commissary's freezers were hacked

Fourteen military commissary freezers across multiple bases simultaneously entered defrost cycles, actively heating food overnight and causing spoilage—an incident that coincided with public disclosure of vulnerabilities in internet-connected refrigeration systems. While the author suspects cyberattack, the technical community largely attributes it to firmware bugs, misconfigured updates, or batch failures. The episode exposes a critical infrastructure design flaw: essential military food storage systems depend on internet connectivity despite no clear operational necessity, creating attack surface and diagnostic ambiguity when failures occur. The incident raises broader questions about defense-in-depth and the sociotechnical risks of connecting non-cyber systems to networks.

Cal Paterson

Memoryfields: A Simpler Approach to Agent Memory

Cal Paterson proposes memoryfields as a minimal, open-standard file format for AI agent memory that treats memory as portable data rather than proprietary pipelines. The approach uses markdown pages with optional YAML metadata and a SQLite vector index, enabling agents to retrieve relevant information through single semantic searches instead of expensive graph traversals. By reducing the machinery between agents and their knowledge stores, the design prioritizes clarity, portability, and graceful scaling with model improvements—a perspective particularly valuable for systems where simplicity directly enables safety and reduces vendor lock-in risks.

LWN.net

Bootstrappable Builds: How and Why

Bootstrappable builds enable creating software from minimal seeds without relying on pre-compiled binaries, directly addressing Ken Thompson's 'trusting trust' attack where malicious code can hide invisibly in compiler binaries. Recent advances like GNU Guix (reduced bootstrap seed from 250MB to 256 bytes) and Live-bootstrap demonstrate practical progress, though complete kernel bootstrapping remains unsolved. Timothy Sample's Germ project proposes a simpler 2.25KB Scheme interpreter approach prioritizing conceptual clarity, though with performance trade-offs. The broader significance for system safety lies in reducing the supply chain verification burden from thousands of steps to manageable numbers, meaningfully constraining the attack surface.

Cloudflare Blog

Introducing Adaptive Intelligence: undermining the economics of every bot attack

Cloudflare introduced Adaptive Intelligence, a machine learning system that continuously retrains on live traffic to detect and defeat bot attacks, replacing static rule-based defenses that attackers can reverse-engineer and bypass. The system uses disposable, non-deterministic detection rules that prevent attackers from isolating exploitable patterns, while deploying new detections in shadow mode before production activation. By shifting from a deterministic model to statistical judgment, Adaptive Intelligence inverts the economic incentive structure—making it prohibitively expensive for attackers to maintain botnets against a defense surface that continuously evolves.