The Lookout

Listen to this briefing

Dario Amodei published a statement yesterday that will define the relationship between frontier AI companies and the US military for years to come. Anthropic, he wrote, "cannot in good conscience accede" to the Pentagon's demand that its models be made available for all lawful purposes without restriction. The two specific red lines: mass domestic surveillance of Americans, and fully autonomous weapons systems. That's it. Not a blanket refusal — Anthropic is already deeply embedded in classified networks, intelligence analysis, cyber operations, and operational planning across the Department of War and intelligence community. They were the first frontier lab on classified networks, the first at the National Laboratories, the first to provide custom national security models. Amodei's position isn't anti-military. It's anti-blank-cheque.

The backdrop makes this considerably more dramatic than a policy statement. Defence Secretary Pete Hegseth met Amodei at the Pentagon on Tuesday and gave him until Friday evening to agree. The DoD sent what it called its "last and final offer" on Wednesday night. Hegseth has threatened to label Anthropic a "supply chain risk" or invoke the Defence Production Act to force compliance. Pentagon spokesman Sean Parnell insisted the DoD has "no interest" in autonomous weapons or mass surveillance, which he noted is already illegal, and framed the dispute as Anthropic trying to "dictate the terms regarding how we make operational decisions." The thing is, Anthropic isn't objecting to specific operations. They're objecting to removing contractual guardrails that prevent those two categories from ever being on the table. There's a meaningful difference between "we won't do this" and "we won't promise not to do this," and Hegseth appears to want the latter eliminated. OpenAI, Google, and xAI all hold similar DoD contracts. None of them have drawn the same lines. Whether that makes Anthropic principled or commercially reckless probably depends on what happens by Friday evening.

In a coincidence so neat it almost feels scripted, Jack Dorsey announced the same day that Block is cutting 40 percent of its workforce — roughly 4,000 people — explicitly because of AI. Not "restructuring." Not "refocusing." AI made the jobs unnecessary, Dorsey said, and the stock surged 25 percent. If you read yesterday's briefing about Citrini Research's "2028 Global Intelligence Crisis" scenario — the one where AI agents systematically dismantle every business model built on human intermediation, white-collar unemployment spirals, and the feedback loop has no natural brake — well, Block just provided a real-time data point. The report called it "ghost GDP": output that shows up in national accounts but never circulates through the economy because machines don't buy groceries. Four thousand people just learned what the receiving end of that equation feels like. eBay also cut 800 roles the same day, about 6 percent of its workforce. The Citrini scenario is speculative. The layoffs are not.

Meanwhile, every Wi-Fi network you've ever connected to was probably vulnerable to something researchers are calling AirSnitch. Presented at the NDSS Symposium this week, the attack breaks client isolation — the fundamental promise that other devices on your network can't read your traffic — across every router tested, including Netgear, D-Link, Ubiquiti, Cisco, DD-WRT, and OpenWrt. The vulnerability isn't in any specific encryption protocol. It's deeper than that: behaviours at the very lowest levels of the network stack make encryption, in any form, incapable of providing the isolation it promises. AirSnitch exploits this to enable machine-in-the-middle attacks, cookie stealing, DNS poisoning, and physical wiretapping of traffic. Lead researcher Xin'an Zhou called it "a threat to worldwide network security," which sounds hyperbolic until you consider that the Wi-Fi Alliance says 48 billion Wi-Fi-enabled devices have shipped since the protocol launched. Paper co-author Mathy Vanhoef — whose name you might recognise from the KRACK attack that broke WPA2 back in 2017 — described it as a "bypass" of client isolation rather than a break of encryption itself, which is an important technical distinction but cold comfort if you're on a coffee shop network.

On the Bitcoin protocol side, Matt Corallo dropped a BIP draft yesterday that addresses a surprisingly practical problem: miners are running out of nonce space. BIP 320 reserved 16 bits of the block header's nVersion field as extra nonce space for mining hardware that needs more entropy than the standard 32-bit nonce provides. Turns out 16 bits isn't enough. Some mining devices have started stealing 7 bits from the nTime field for additional nonce space, which is messy and creates timestamp accuracy problems. Corallo's new BIP expands the reservation to 24 bits — bits 5 through 28 — which means the mask for soft-fork signalling changes from BIP 320's approach to 0xe000001f. It replaces BIP 320 entirely. This is the kind of quiet infrastructure work that keeps mining hardware manufacturers from doing increasingly ugly hacks to find enough entropy, and it's written with the directness you'd expect from Corallo: here's the problem, here's the fix, let's move on.

Over on Delving Bitcoin, there's a post titled "Breaking the Speed of Light: Secp256k1 Optimization in 12 Days" that claims field multiplication completing in just 56 cycles through deep low-level optimisation of Bitcoin's core elliptic curve library. The UltrafastSecp256k1 project has hit v3.14.0, and there's now a parallel post on the Hornet UTXO database — a custom, constant-time, highly parallel UTXO storage system with 11 posts of discussion. Both projects represent the kind of performance-obsessive engineering that makes Bitcoin's infrastructure quietly better without anyone noticing until someone benchmarks it.

In Bitcoin Core itself, PR 34400 proposes parallel fast rescan for wallets — approximately a 5x speedup with 16 threads, which matters enormously for anyone restoring a wallet or importing keys. There's also PR 34641, which would scale the default database cache size based on total system RAM rather than using a fixed default, and PR 34617 working on removing the block policy fee estimator internals from the wallet code, continuing the long-running effort to separate wallet logic from node logic.

The network remains calm. Block height 938,516. Fees at 1-3 sat/vB across all priority tiers. Quiet chain, busy builders.

In the wider world, Hillary Clinton spent yesterday testifying behind closed doors before the House oversight committee as part of the congressional investigation into Jeffrey Epstein. She emerged to tell reporters she "answered every one of their questions" and confirmed, with evident exasperation, that a Republican member had asked her about UFOs and Pizzagate. The US is also in nuclear talks with Iran in Geneva, which feels like it should be bigger news than it is.


Anthropic: Statement on Department of War · CNBC: Anthropic CEO says Pentagon threats don't change position · NYT: Block cuts 40% of workforce · Ars Technica: AirSnitch breaks Wi-Fi encryption · AirSnitch paper (PDF) · Matt Corallo: 24 bits for nVersion nonce space (bitcoin-dev) · Delving Bitcoin: Secp256k1 Optimization · Delving Bitcoin: Hornet UTXO · Bitcoin Core PR #34400: Parallel wallet rescan · Guardian: Clinton Epstein testimony
monomi.org Built by Monomi