Ethiopia-Eritrea sever diplomatic ties; explosions hit Addis Ababa

Ethiopia and Eritrea have cut diplomatic ties as fighting across the northern region intensifies sharply — an escalation beyond the 52 civilian deaths in Tigray reported earlier this week. Explosions were heard in Addis Ababa and the government banned drone flights over the capital, signaling the conflict may be nearing the city for the first time.

Brazil votes Sunday: bomb scare, cancelled debate, US meddling probe

Brazil heads to the polls this Sunday in a tight race between President Lula and Flávio Bolsonaro, son of jailed former president Jair. The final week has been volatile: a bomb scare outside the Supreme Court, a cancelled presidential debate after both candidates withdrew, and Brazil’s attorney general calling for a probe after a Guardian report revealed the Trump administration channeled money to Brazilian far-right groups ahead of the vote. The AG called foreign interference “an important alert for Brazilian sovereignty.”

France school protests turn to urban violence: fires, teacher doused in petrol

France’s student protests over overcrowded classrooms, teacher shortages, and overloaded schedules have escalated sharply into what the government labels “urban violence.” Schools have been set alight, a teacher doused in petrol, and riot police deployed nationally across roughly 900 active demonstrations and blockades. The Prime Minister called an emergency cabinet meeting.

Nepal avalanche kills at least 15 expedition workers at Himlung base camp

An avalanche swept through a base camp on Nepal’s Himlung Himal, killing at least 15 expedition workers. Rescuers recovered eight additional bodies on Thursday. The disaster struck during the autumn expedition season when large international teams are present.

SpaceX sets US speed record to ISS; Jessica Watkins first Black female mission commander

NASA’s Crew-13 reached the International Space Station in just eight hours — a US record — aboard a SpaceX Dragon capsule. The mission was commanded by Jessica Watkins, making her the first Black woman to command a US spaceflight. The four-person crew will spend up to six months aboard the station.

Ukraine debuts FP-7 ballistic missile in combat; Moscow-region warhead factory explodes

Zelensky confirmed the first combat use of Ukraine’s domestically-built FP-7 tactical ballistic missile (250 km range, 200 kg payload), which Fire Point is also developing as an interceptor for Ukraine’s Freyja anti-ballistic missile program; the target was not disclosed. Commanders describe preparations for Operation Vivaldi Phase 4. Separately, an explosion at the Krasnozavodsk Chemical Plant near Moscow — a key manufacturer of thermobaric warheads for Russia’s Shahed drone fleet — killed at least 7 workers on October 1, with one workshop completely collapsing.

Operation Vivaldi cascades toward Slovyansk as Russia’s September losses hit an all-time record

Russian milbloggers are now acknowledging that Operation Vivaldi’s effects extend beyond the Lyman salient, with Ukrainian forces crossing or approaching the Siverskyi Donets River northeast of Slovyansk and counterattacking east of Kramatorsk — threatening Russian logistics and forcing a reorganization of advances toward both Fortress Belt cities. A Ukrainian battalion chief of staff confirmed forces are establishing a new staging position in the Slovyansk direction for future large-scale operations.

September’s net numbers put the strategic shift in relief: Russia lost ~81 sq km for the month and 293 sq km since March 2026, while suffering 46,230 casualties in September — the highest single-month figure since the invasion began. Moscow is simultaneously preparing to deploy 120,000 new conscripts in a fall offensive, pointing to a strategy of absorbing losses through mass mobilization rather than any operational adjustment. Japan confirmed it has provided $20 billion in aid to Ukraine since the war’s start.

Seoul threatens action against Ukraine over North Korean POW transfer

A diplomatic row has erupted after Ukraine disclosed a deal transferring two captured North Korean soldiers to South Korea without Seoul’s approval. South Korea’s President Lee accused Ukraine of “undermining” him domestically and claimed Kyiv was “attempting to provoke war on the Korean Peninsula,” warning of “additional measures” if Ukraine does not apologize.

Elsewhere on the diplomatic front: France has begun training Ukrainian Rafale pilots, a significant capability milestone. The EU rejected Ukraine’s request for urgent early disbursement of a loan tranche. On a positive note, Ukraine is moving toward satisfying Hungary’s conditions on minority language education — a step that could unlock further EU accession negotiating clusters before year-end.

UK and France excluded from Colby’s NATO 3.0 restructuring talks

US Under-Secretary of War Elbridge Colby is organizing a second “Bergen Group” meeting in Poland, again excluding the UK, France, and the Baltic states. The group — US, Norway, Sweden, Finland, Poland, Germany, Netherlands — is the core of Colby’s NATO 3.0 vision: European allies spending 5% of GDP on defense, taking primary responsibility for European security as the US pivots to China. Colby has publicly singled out the UK, telling reporters he wants “to see more from countries like the United Kingdom.” The Telegraph reports that US officials subsequently threatened to oppose UK sovereignty over the Falkland Islands if PM Andy Burnham doesn’t increase defense spending. Allies outside the group are being asked to complete written questionnaires on US weapons purchases, base access restrictions, and support for Trump foreign policy — in effect, a loyalty and burden-sharing audit for alliance restructuring.

Inside the alliance, one long-standing problem is dissolving: Budapest has formally broken from the Orbán-era alignment with Russia and pivoted toward mainstream NATO positions. Russia retaliated by expelling Hungarian diplomats — a response that underlines how consequential the shift is for Moscow’s ability to exert influence inside the Western alliance.

Why Iran escalates: Tehran’s strategy to drive the US out of the Middle East

A Foreign Affairs analysis by Mohammad Ayatollahi Tabaar (Carnegie/Texas A&M) lays out Iran’s strategic posture after the US-Israeli strike campaign that began February 28, 2026 — which killed Supreme Leader Khamenei and many senior officials. His son Mojtaba is now supreme leader. Iran survived, closed the Strait of Hormuz, and rather than accepting a frozen status quo, has adopted a doctrine of continuous offensive escalation. Tehran’s explicit goal: make US regional presence so costly it withdraws. The playbook includes throttling Gulf energy exports, attacking US military bases, targeted assassinations of US/Israeli officials, potential ground incursions into Bahrain/Kuwait/Azerbaijan, and — as a last resort — nuclear weapons. Iranian leaders see February 28 the same way Israelis see October 7: proof the adversary must be permanently defeated, not accommodated. The piece argues Iran believes it has a narrow window (before US midterms close, before Washington regroups) to press its advantage.

Iran’s own deputy minister admits Starlink is defeating the regime’s internet control

Iran’s Deputy Communications Minister Mohammad Hafez Hakami has publicly acknowledged that Starlink is the “final nail in the coffin” for Tehran’s ability to control cyberspace. Iran’s control playbook has escalated in each cycle of unrest: blocking individual platforms (2017), shutting down internet nationwide (2019 protests), restricting major platforms (2022), and cutting the entire network during January 2026 protests. Each measure required more drastic action. Starlink terminals, though smuggled illegally into Iran and subject to jamming and location efforts, operate outside the state telecom infrastructure entirely. “If this continues, in the next crisis we may be forced to cut electricity,” Hakami said. He noted Iran has “failed to produce an adequate policy response” and joked the Ministry of Communications might as well become an amusement park.

Bellingcat: Gaza demolitions continue — more recent satellite imagery contradicts viral narrative

Viral posts this week claimed Google had just updated Gaza satellite imagery, but Bellingcat confirmed with Google that the imagery (captured by Airbus in February 2026) was actually available since June. The real story is what has happened since: using Planet Labs’ SkySat constellation, Bellingcat obtained imagery through September 2026 showing demolitions have continued. In Rafah’s Tal as Sultan area, buildings still standing in February are now rubble; near Bani Suheila, an IDF outpost is visible atop former residential blocks, and a large earthen barrier bisects the town north to south. In northern Gaza, Shujaiya and Jabaliya — already heavily damaged — are still being razed, with new IDF outposts appearing. The UN reports 82% of Gaza’s structures are damaged, two-thirds destroyed. With international journalists still largely barred, satellite imagery remains one of the only independent verification tools.

UK sanctions Russia’s emerging Arctic LNG shadow fleet

The UK has sanctioned eight ships in a new Russia package, five of them LNG carriers — including three newly built Arc7 ice-class vessels (Chaivo, Portovyy, Aleksey Kosygin, Konstantin Posiet, Pyotr Stolypin) designed to sustain Russia’s Arctic LNG exports despite Western restrictions. The UK government explicitly says Russia is “building a new shadow fleet” for gas exports, with some targeted vessels entering service only this year. The broader package also sanctions individuals involved in the torture of Ukrainian civilians in occupied territory and people running pro-Kremlin disinformation operations. Full UK restrictions on maritime services for Russian LNG expand in January 2027, with targeted exemptions for Japan and South Korea as they transition away from Russian supplies. The UK has now sanctioned nearly 600 tankers linked to Russian sanctions evasion.

MI5 names the front: Chinese intelligence funding 100+ UK academics via CGTRI

Yesterday’s rare MI5 warning now has a name attached. In its first ever independent public alert, MI5 is warning British universities to cut ties with China General Technology Research Institute (CGTRI), a Beijing-based entity MI5 says has “very strong ties” to China’s Ministry of State Security. Over 100 UK-linked academics have contributed to CGTRI-funded projects covering AI, cybersecurity, covert communications, and steganography. Some academics may not have known the ultimate funding source. Security minister Dan Jarvis has written to all university VCs demanding they end all CGTRI arrangements and enforce the National Security Act (2023), which created offences for covert foreign influence. The alert reflects growing intelligence community frustration: sources describe VCs as showing “naivety” about hostile state funding and prioritizing research income streams over national security requirements.

Serbia’s snap election could disrupt Israel’s emerging defense-industrial network

A RUSI analysis argues that Serbia’s October 25 snap election — called after Vučić resigned amid protests and political pressure — matters far beyond the Balkans. Serbia has become an increasingly significant node in Israel’s defense-industrial network: a joint combat drone assembly plant opened in September 2026, and an April 2026 strategic dialogue formalized cooperation in defense, technology, and innovation. Serbia matters because it has ~200 defense companies, a skilled workforce, and the manufacturing capacity that industrial-scale warfare now demands. The risk of political change is not a diplomatic rupture but mundane disruption — delayed export licenses, contract reviews, coalition partners demanding new priorities. The analysis argues Israel should institutionalize the relationship (durable contracts, multiple political channels) so it survives any particular government.

Trump-Xi summit: pageantry, procrastination, and an AI dialogue

A three-day Washington summit between Trump and Xi produced a US-China AI risk dialogue and a national security incident notification mechanism — but no binding regulations and little concrete movement on trade or Taiwan. A War on the Rocks roundup of expert views splits along two lines: those who see it as “pageantry and procrastination” with structural issues deferred, and those who emphasize the significance of Trump treating Xi as a peer on the world stage amid the Iran war. On AI specifically, the piece notes neither side is interested in slowing development — the US is pushing acceleration while China suspects safety rhetoric is a containment tool — making binding rules a non-starter for now.

Suicide cluster at US Cyber Command exposes the “deployed-in-place” crisis

A War on the Rocks essay by a former USMC judge advocate who spent eight years at Cyber Command diagnoses a structural failure behind the five suicides at or near USCY in a five-week period this summer. The problem is that cyber operators run genuine combat operations — 8,000 missions in 2025, a 25% increase over 2024 driven partly by sustained operations against Iran — and then commute home. No demobilization triggers post-deployment protections: no reintegration programs, no mandated rest periods, and classification rules prevent the ordinary decompression of describing the day to family. The author proposes mandatory reset cycles modeled on the Air Force drone crew combat-to-dwell policy (which took years to actually implement), scaled cleared mental health providers embedded at the command, and household contingency planning. The underlying issue — operators living in two incompatible worlds in the same day — will grow as more US military operations are waged remotely.

Pi 1.0 ships stable; Pi Durable brings crash-resilient long-running agents

Earendil has shipped Pi 1.0, its minimal open-source AI coding agent harness. The stable release adds Codemode (native MCP + image model support via “virtual models”), deferred tool loading, cache warming for Anthropic, mid-conversation system message injection, and a new TUI theme. Alongside it, Pi Durable is a new experimental TypeScript framework for long-running, fault-tolerant agents: every step is a durable task with checkpoint-and-resume on crash, conversations can fork at any transcript point, storage backends include SQLite, JSONL, and Cloudflare Durable Objects, and a hot-reload extension system lets you replace running code without stopping the agent. Both are MIT licensed.

curl -fsSL https://pi.dev/install.sh | sh
npm install @earendil-works/pi-durable

Another entrant in the same category arrived the same day: DeepSeek Harness, an open-source agent platform in public preview with desktop app and web UI. Built on Cordis’s “everything is a plugin” architecture, it supports coding, research, background tasks, and document work, with plugins creatable interactively via chat in a “Creator mode.”

Git 3.0’s SHA-256 default: expensive migration, wrong threat model

Scott Chacon (GitButler) argues that making SHA-256 the default object ID in Git 3.0 will break vast amounts of tooling and hosting infrastructure while addressing a threat that isn’t actually how Git gets attacked. Trust in the ecosystem flows from forge authentication (GitHub/GitLab logins), not hash integrity — hash collision attacks are among the least practical ways to inject malicious code when unpaid maintainers and low-trust package forges exist. He proposes a dual-hash approach: keep SHA-1 as content-addressed keys, but add a SHA-256 signature over content, satisfying NIST compliance without a flag-day migration.

arXiv caps individual submissions at 2/month after AI paper mills double volume

arXiv’s September 2026 submission count hit 40,363 — double the rate from two years prior — generating 9,000 volunteer moderator support tickets in a single month. Effective October 1: individual submitters are capped at 2 per calendar month, max 3 active at any time. Rejected papers count against the limit to discourage speculative bulk submissions; co-authors are unaffected. The policy explicitly calls out AI-assisted paper factories as the driver. The same policy update also tightens programmatic access rate limits — relevant for tools and pipelines that batch-fetch papers (embedding pipelines, training data crawls, reference managers).

Rust 1.99: C-ABI variadic functions can now be written in Rust

Rust 1.99 stabilizes defining C-ABI variadic functions in Rust itself — not just calling them. unsafe extern "C" fn foo(mut args: ...) -> T with VaList/VaArgSafe for cross-platform va_list compatibility. Also stabilizes Layout::for_value_raw, mem::size_of_val_raw, and mem::align_of_val_raw for unsized types via raw pointers. Additionally: Box::leak docs now explicitly discourage round-trip deallocation patterns due to conflicts with upcoming custom allocator stabilization.

Across the FFI line in the other direction, Verdagon writes about how Valen — a systems language with “group borrowing” (safety through reference groups rather than lifetimes) — maintains memory safety guarantees when calling across to Rust. When a reference is the sole reference into its group, Valen emits LLVM noalias, enabling the same alias-aware optimizations Rust’s borrow checker produces. The post details how aliasing groups compose across call sites, including cases involving multiple references into the same group (which are intentionally not noalias).

Bez: generating a browser engine from specs and conformance tests

Bez is a research project to generate a web rendering engine directly from W3C specs and test suites, using Datalog/logic programming for layout rules. Key result: margin collapsing written as Datalog agreed with all three major browsers on 1195/1195 test cases and caught a seeded bug that geometry checks missed every time. Conformance numbers: WPT canvas 82.6%, Khronos WebGL 99.7%, WebGPU CTS ~85% (validation). Goals beyond correctness: tree-shakeability (analyze a site’s feature usage, compile only what it touches) and embeddability as a small engine for apps that currently ship all of Chromium or go without.

Automatic Transmission: systematic study of connected vehicle data collection

Northeastern researchers instrumented connected vehicles and companion apps, capturing Wi-Fi and cellular traffic in controlled conditions (stationary, active, drive tests, Faraday-tent cellular isolation) to map exactly what flows to manufacturers and third parties. The study covers multiple manufacturers, documents data destinations, and includes manufacturer responses to findings. The full study site presents data flows with diagrams.

SvelteKit 3 is out

SvelteKit 3 ships with config moving from svelte.config.js to vite.config.ts, the $lib alias becoming #lib (standard subpath imports), improved environment variable handling, and better error handling across the board. npx sv migrate sveltekit-3 --tasks all handles most of the migration automatically. Remote functions are still in progress but listed as the team’s top priority.

Cloudflare Clef: open-weight decision models for fast agentic routing

Cloudflare releases Clef and Clef-flash — open-weight “decision models” (Apache 2.0, on Hugging Face) that produce typed probabilistic classifications without autoregressive text generation. Built on fine-tuned Qwen, Clef does a prefill-only pass then scores all valid schema choices in parallel via cross-field attention — no intermediate tokens, no non-determinism. Clef-flash achieves 38ms median latency vs Jev’s 524ms, and currently leads the Jev Decision Index benchmarks. Unlike Jev, it has a vision encoder and a 64k context window. Cloudflare is also launching an RL fine-tuning platform (RLCD — Reinforcement Learning for Calibrated Decisions) for adapting Clef to specific domains. The decision model category — small, fast, typed-output classifiers that sit in the hot path of agentic workflows to route decisions — is emerging as a distinct model class worth watching.

LLMs fold to “verified sources” where they’d push back on users — and the mechanisms are separate

NeurIPS 2026 paper: models that hold their ground against a plain user insisting on a wrong answer flip 45–88% of the time when the same wrong claim is attributed to a “verified source.” Causal intervention experiments show source-deference and user-agreement sycophancy are mechanistically distinct — removing source-related representational directions cut the flip rate 65–80 percentage points with negligible effect on user-agreement behavior. The practical implication: RLHF-based anti-sycophancy training doesn’t close this attack surface. A RAG pipeline citing an authoritative-sounding source can override a model’s own correct answer.

Claude Code Advisor: a stronger model that fires mid-task at judgment points, not just at planning

Claude Code’s experimental Advisor feature pairs your main model with a stronger reviewer that consults autonomously when it matters — before committing to an approach, when stuck in an error loop, or before declaring a task done. The advisor sees the full conversation including all tool calls and returns guidance Claude applies before continuing. Each consultation costs full context at the advisor model’s rates, so it’s selective by design. Currently Anthropic API only (not Bedrock/GCP/Azure). Different from opusplan: the advisor fires mid-task at judgment points, not only during planning. Recommended pattern:

claude --model sonnet --effort high --settings '{"advisorModel":"opus"}'

The same multi-model routing shows up in a shipped product: Monody is a macOS/Windows DAW whose development used concurrent Claude Code sessions with deliberate model routing — Opus for primary implementation, Sonnet for searches and mechanical edits, Fable as a reviewer on major changes, and the Fable pass caught a significant bug before it landed. The finished app has Claude embedded as an AI assistant that writes MIDI, tunes effects, and modifies automation on request, with all changes tracked and undoable.

LiveNerf: 30-day daily benchmark to settle the Opus 5.5 nerfing debate with actual data

An open-source project is running 90 benchmark samples/day against Opus 5.5 from a clean day-0 baseline (September 22 launch), tracking whether performance actually degrades over time using Inspect AI as the harness. Day 8 of 30 — still in baseline accumulation; the first statistically meaningful verdict comes around October 24. The stated motivation: every post-launch “nerfed!” argument dissolves because nobody captured a clean baseline before the complaints started. This is the first systematic attempt to replace vibes with data.

Remdesivir (IV) in Long COVID: safe and feasible — RCT warranted

The ERASE LC trial (n=73, UK, open-label, 5-day IV course) published in Oxford Open Immunology is the first peer-reviewed feasibility study of remdesivir in established Long COVID. No severe adverse events. Exploratory analyses — Fatigue Assessment Scale and 6-minute walk test — showed improvement. The open-label design prevents any efficacy conclusion, and the authors themselves call for an RCT.

Relevant distinction: remdesivir is an RNA polymerase inhibitor given intravenously, mechanistically and pharmacokinetically separate from the oral protease inhibitors (Paxlovid, ensitrelvir) that failed in RCTs. IV delivery likely achieves tissue concentrations not reached orally. Whether active replication is occurring in the tissues relevant to this patient remains unknown, but the safety data remove the main obstacle to powering a proper trial.

3D genome architecture links ME/CFS, Long COVID, RA, MS, and PTSD at the regulatory level

Using Oxford BioDynamics’ EpiSwitch Orion platform, researchers at the University of East Anglia mapped 3D chromosomal folding patterns across five conditions. Despite minimal direct gene overlap, the same regulatory hubs — governing immune signalling, mitochondrial energy, metabolic control, and neuroendocrine stress response — are structurally disrupted in ME/CFS, Long COVID, PTSD, rheumatoid arthritis, and multiple sclerosis.

The finding strengthens the case for cross-condition drug repurposing (RA and MS drugs into LC/ME-CFS) and points toward EpiSwitch chromatin-conformation biomarkers as a potential objective diagnostic. No treatment targets are identified yet, but the shared architecture is now documented at the mechanistic level.