OpenAI ships the GPT-5.6 family right as its own safety testers find universal jailbreaks; the EU extends warrantless mass chat-scanning through 2028 despite Parliament rejecting it; the AI Futures Project publishes Plan A, a full US-China roadmap to safe superintelligence

Patrick Breyer

EU Parliament Extends Mass Chat-Scanning Through 2028 Despite Rejecting It

The European Parliament voted to extend 'Chat Control 1.0' — the temporary derogation that lets platforms like Instagram, Discord, and Gmail voluntarily scan private messages for CSAM without a warrant or individualized suspicion — through 2028, even though 314 MEPs voted against the underlying approach. MEP and digital-rights advocate Patrick Breyer called the outcome 'a farce that damages democracy': the vote didn't ratify indiscriminate scanning as good policy, it just punted the harder political fight — over whether the permanent successor law should target only actual suspects or continue scanning everyone's messages by default — to negotiations resuming in September. For a technical audience, the interesting failure mode isn't the politics so much as the pattern: temporary, ostensibly voluntary security measures adopted under emergency framing have a well-documented tendency to become permanent infrastructure precisely because sunset clauses create renewal decisions that are procedurally easier to win than the up-front fight ever was.

alexalejandre.com

Mitchell Hashimoto on Ghostty, Zig, and Why Open Source Owes You Nothing

In a wide-ranging interview, Ghostty and Vouch creator Mitchell Hashimoto lays out a maintainer philosophy built around a blunt reading of open-source licenses: 'as-is, no warranty' means no entitlement to demands, and the honest response to 'this doesn't do what I want' is 'then fork it' — not as a dismissal, but as a sincere invitation, since forking is the actual mechanism the license grants you instead of a support contract. That same distaste for unearned obligation shapes his terminal design: Ghostty's guiding architectural rule is that a feature should cost you nothing — no runtime overhead, no cognitive tax — unless you actually use it, which is how he squares 'feature-rich' with an aversion to bloat. He's also sketching genuinely new terminal protocols (an 'n-screen' API to replace the primary/alternate-screen binary, a button protocol extending OSC 8 hyperlinks to scrollback-persistent interactive elements) at a moment when, he argues, terminals have had no real standards body for twenty years. The aside on AI is worth noting for anyone tracking how frontier coding tools change engineering judgment, not just throughput: he describes using AI to draft huge, otherwise-prohibitively-expensive backward-compatibility migrations across Zig's changing APIs, arguing it structurally lowers the cost of breaking changes — a real shift in what 'stability' is worth engineering for in the first place.

GitHub Blog

How GitHub Gave Every Repository a Durable Owner

GitHub's internal security team turned a nagging operational problem — over 11,000 of its roughly 14,000 internal repositories had no reliably identifiable owner — into a case study on retrofitting accountability onto systems that grew organically. The fix wasn't a new tool so much as making ownership a first-class, enforced property: two GitHub custom properties (ownership-type, ownership-name) validated against real team membership and a live service catalog, synced automatically where possible, then enforced with a 30-day grace period and reversible archival for anything left unclaimed. What makes the writeup useful beyond the specific mechanism is the postmortem honesty about the two incidents that shaped the final design: archiving a repo silently broke a Datadog integration that paged an on-call team with no context, so they added mandatory notifications to repo admins; and a hypothetical stale Service Catalog read could have mass-archived legitimate repositories, so they added a circuit-breaker 'low water mark' that halts the job entirely if it's about to take an anomalously large action. Neither safeguard was in the original design — both came from the system actually breaking once — which is the recurring lesson in build-outs like this: the guardrails worth having are usually the ones you can only discover by getting burned first, so the goal is designing the blast radius small enough that getting burned is cheap.

Modern War Institute at West Point

The Glass Backbone: Why the Army's Logistics Will Break in the Next War

A Modern War Institute analysis argues that the U.S. Army's logistics architecture — built and refined over two decades of counterinsurgency, where rear-area supply hubs were, by definition, safe from enemy fire — is structurally unsurvivable against a peer adversary with persistent satellite surveillance and long-range precision strike. The argument is essentially a systems-safety one dressed in military language: centralizing throughput into a small number of large, fixed, easily-targeted nodes is efficient precisely because it removes the redundancy that would let the network survive losing any single node, and Ukraine has already demonstrated in practice what happens when that tradeoff meets modern targeting — concentrated logistics hubs get found and destroyed. The proposed fix — dispersed, mobile, redundant supply networks using autonomous resupply platforms, rather than fewer-and-bigger depots — is a direct application of the same centralization-versus-resilience tradeoff that shows up in distributed systems design, power grids, and supply chains generally: the efficient topology and the survivable topology are usually different topologies, and organizations that optimize hard for one under peacetime pressure often only discover how exposed that made them once the failure mode they weren't designed for actually shows up.