Unprecedented rains kill at least six in Japan, strand 7,000 at Narita

“Unprecedented” heavy rainfall struck Chiba prefecture in eastern Japan during the busy Bon holiday week, killing at least six people, cutting power to more than 20,000 households, and leaving nearly 7,000 travellers stranded at Tokyo’s Narita airport. Evacuation orders went out to around 400,000 residents, and flooding risks persisted even as the rain began to ease.

Romania forced to shut its only nuclear plant as the Danube falls to record lows

Romania has shut down all reactors at its Cernavodă nuclear power plant — which supplies 20% of the country’s electricity — after a severe drought dropped the Danube to record lows, cutting the water supply needed for cooling. The plant is not expected to restart for at least ten days. The shutdown creates an immediate crisis for Moldova, which relies on Romanian power to cover its own chronic electricity shortages.

UK hits 38.1°C — hottest day of 2026 — as Greece battles wildfires

Britain recorded its hottest day of the year at 38.1°C as the severe European heatwave continued, with a former government scientific adviser warning that “worse is still to come.” Greece simultaneously fought wildfires fanned by gale-force winds. One index of how the expectation is settling in: Rightmove reports searches for UK homes with air conditioning have more than doubled this summer.

Oil slick off Oman expands to 1,300 sq km as salvage begins

The slick from the Caroline Bezengi — a tanker carrying sanctioned Russian crude that ran aground off Oman — has expanded to cover 1,300 square kilometres according to Greenpeace, threatening fragile Gulf ecosystems. Salvage operations have now begun on the half-submerged vessel. The spill has grown dramatically since first reaching Oman’s coastline days ago.

US government ran mass surveillance campaign on leftwing groups and anti-ICE protesters

Newly disclosed records reveal DHS deployed undercover agents to infiltrate meetings and private chats of prominent leftwing organisations and groups opposing Trump’s immigration crackdown — one of the largest known domestic intelligence campaigns against civil society groups in recent years, and well beyond what had previously been publicly acknowledged.

Flock Safety adds guardrails after police used its plate network to target immigrants and abortion seekers

Flock Safety — whose system conducts roughly 20 billion licence-plate scans annually across the US — is introducing new safeguards after reports that police used the technology to track undocumented immigrants and people seeking out-of-state abortions, and after a wave of stalking cases enabled by the same infrastructure. The CEO announced oversight steps; critics say the problem is structural.

US accuses 40+ nations of helping China dodge tariffs; Brazil opens retaliation process

A new White House report accuses more than 40 countries of helping China route goods through lower-tariff nations to evade Trump’s levies — the most explicit public accounting yet of the scale of circumvention. Brazil separately opened a formal reciprocity process against the US over its own tariffs, exploring retaliatory measures without committing to them.

Foreign Affairs runs an argument that sits directly underneath this: Michael Froman, the former US Trade Representative, makes the case that China’s structural overcapacity, deflationary export pressure and debt accumulation are the most likely source of the next global economic crisis — and that the world is poorly positioned to absorb the shock given existing strain from the Iran conflict and post-pandemic debt loads. Unlike previous crises, he argues, the transmission would be faster and more synchronised, running through trade channels.

Medvedev threatens Japan with ‘monstrous consequences’ over the Kurils

Russia escalated its rhetoric after Putin’s first-ever visit to the disputed Kuril Islands, with former president Medvedev threatening Japan with “monstrous consequences” following Tokyo’s sharp protests. Japan’s PM had called the visit “absolutely unacceptable,” asserting the islands — seized by the Soviets in the closing days of the Second World War — belong to Japan under international law. The two countries have never signed a peace treaty.

New carrier heads to the Mideast as sailors on the USS Abraham Lincoln face a 250-day breakdown in conditions

The USS George Washington has deployed toward the Middle East as alarming reports emerge from the USS Abraham Lincoln, at sea for more than 250 days. Sailors on the Lincoln have reportedly endured food shortages, broken plumbing and deteriorating mental health, with some said to have considered jumping overboard. The extended deployment is a direct consequence of the sustained US commitment to the Iran war.

Farage retains Clacton in a boycotted by-election he triggered himself

Nigel Farage won the Clacton by-election — a contest he caused by resigning his seat, boycotted by every major party, and which he did not attend to receive the result — defeating nearest challenger Count Binface. The vote arrives amid deepening scandal over a £5 million gift to Reform UK and an ongoing donations investigation. Farage said the result “speaks for itself,” without appearing at the count.

Deep strikes push further: Ust-Luga, Moscow, Ufa — and the fuel crisis bites

Fifty-four Ukrainian drones set the Ust-Luga petroleum export terminal ablaze overnight — Russia’s largest Baltic oil export port, terminus of the Baltic Pipeline System-2, and 110 km from St Petersburg. Drones also reached Moscow. A Wildberries logistics hub near Ufa in Bashkortostan, 1,300 km from the front, was hit for the second time with AN-196 Liutyi drones.

The cumulative effect of the campaign is now visible inside Russia rather than only on the target list: seaborne oil product exports fell 55% year-over-year in July, the Orsk refinery is offline for up to six months, and fuel queues and rationing have returned across multiple regions.

Ukraine offers a Black Sea truce; Zelenskyy rules out land concessions

With food supply fears mounting, Ukraine has offered Russia a ceasefire in the Black Sea, even as its drone strikes continue to upend daily life in Crimea. Zelenskyy warned that any territorial concession would only give Putin time to prepare another invasion, and disclosed that Ukraine needs hundreds of Patriot missiles to get through the coming winter.

Vance asked Ukraine to spare the Caspian pipeline; Putin threatens Western ships

The Financial Times reports that VP Vance asked Ukraine in late July not to target Caspian Pipeline Consortium infrastructure or non-sanctioned vessels, and that Kyiv agreed — Washington managing global energy market risk while Ukraine continues to strike everything else. Putin, meanwhile, threatened this week to seize Western shipping, saying Russia tracks where NATO vessels sail: a signal aimed at the same maritime pressure points from the other direction.

UN: July civilian toll the highest since May 2022 — 437 killed

UN monitors recorded 437 civilians killed and 2,610 injured in Ukraine in July, up 30% from June and 70% from July 2025 — the highest monthly death toll since the opening phase of the invasion. Guided aerial bomb deaths jumped 143% from June (105 killed, 753 injured), and short-range FPV drone strikes near the front line killed 111 civilians, their worst monthly figure of the war. Child casualties hit their highest level since April 2022. A Russian double-tap UAV strike today killed a mother and child in Sumy Oblast.

Defector submarine commander killed by bomb in Sevastopol park

Robert Shageev — former commander of Ukraine’s only submarine, who defected to Russia in 2014 and later took part in Russian submarine missile strikes against Ukraine — was killed by a bomb planted in a Sevastopol park bin and detonated remotely as he passed. Russian authorities arrested a woman, saying she acted “on instructions from Ukrainian intelligence.” Five further Russian servicemen died in a secondary explosion while trying to clear the scene — a double-trap. Ukraine has not commented.

Iran learned to exhaust US interceptor stockpiles — then one missile got through

Over five days in July, Iran launched waves of drones and manoeuvring missiles at three US bases in Jordan, deliberately structured to force Americans to burn through scarce Patriot interceptors. On day five, a missile struck prefabricated housing, killing three US soldiers and wounding more than 100.

Pentagon officials have since acknowledged the US interceptor supply is critically depleted. The episode is of a piece with Iran’s broader tactical learning across the conflict — the move from mass-casualty strikes to selective targeting of radars, tankers and SATCOM that the Hudson report ties to Chinese satellite intelligence.

Two intelligence gaps on Iran: an unverified plot and an unexpected rebuild

Reuters reports that US intelligence agencies were unable to corroborate Israeli warnings that Iran was plotting to assassinate Donald Trump — a significant gap, given how much escalatory decision-making during an active conflict has rested on Israeli assessments, and given Israel’s own interest in sustaining US pressure on Tehran.

Running the other way, Israeli intelligence has reportedly been caught off guard by the speed at which Iran is reconstituting its ballistic missile arsenal after the war. The pace is raising alarm about Iran’s long-term regional strike capacity — and it arrives precisely as Washington announces an indefinite blockade and new economic isolation measures.

West Bank: IDF forces families from their homes as the settler rampage reaches a Christian town

Israeli troops evicted two Palestinian families from their besieged homes in Qusra and took the houses over as military barracks — a day after an IDF commander personally promised to end the settler violence. Taybeh, the last all-Christian village in the West Bank, has also come under coordinated settler attack. Israel’s defence minister praised settlers for “holding onto” Palestinian land.

The condemnation has been unusually direct. The US ambassador called the siege a “horrific act of terror” and the administration is now formally on record using that language; the UN called it “criminal and inhumane”; and former Israeli PM Ehud Olmert described it as a “concerted and meticulous attempt at ethnic cleansing.”

The Mecca Pact: Turkey, Saudi Arabia and Pakistan formalise a new defence alliance

Turkey, Saudi Arabia and Pakistan signed the “Mecca Joint Defence Agreement” on 7 August, with an Article 5-equivalent mutual defence clause. Ankara confirmed Thursday it will establish joint coordination mechanisms at defence minister, foreign minister and military chief levels, with joint exercises and defence industry cooperation to follow.

The configuration is novel: nuclear-armed Pakistan takes nominal responsibility for Saudi defence, while Turkey remains a NATO member. If Houthi or Iranian attacks spread to Saudi territory, Ankara could find itself holding overlapping obligations in two alliances at once.

How a bus driver’s kick may have averted a European crisis

An airport bus driver at Leipzig/Halle kicked an explosive-laden drone out of the air before it reached a Ukrainian An-124 heavy transport — one of only 57 worldwide, and central to NATO’s Strategic Airlift International Solution. US intelligence reportedly suspects Russian military intelligence; German federal prosecutors have taken over. A second drone struck another aircraft at Leipzig the same day.

Gen. Mark Hertling’s Bulwark analysis reconstructs the escalation web that would have followed had the driver been seconds slower: a detonated aircraft killing German citizens, NATO paralysed by Article 5 ambiguity over a covert hybrid attack, Russian forces probing the resulting hesitation. He connects Leipzig to the Turkey–Pakistan–Saudi pact, the Iran conflict and potential Chinese moves on Taiwan, arguing that manufactured ambiguity is hybrid warfare’s core feature rather than a side effect — a warehouse fire or a drone under a wing forces governments to establish what happened, who did it, and whether it crosses a threshold before they can respond at all.

Iraq is running out of money

DW reports that the Iran war has strangled Iraq’s economy — disrupting trade, deterring investment and cutting oil-related revenue — leaving the government struggling to pay salaries and fund basic services. Iraq sits at the intersection of the conflict: hosting US bases that came under attack, deeply embedded in Iran’s economic orbit, and politically unable to side fully with either party. The fiscal crisis compounds an already fragile politics dominated by Iranian-aligned factions.

extern “fil-c”: memory-safe C FFI in Rust, with Nix already ahead

Domen Kožar (Cachix, NixOS) proposes an extern "fil-c" FFI mode for Rust. Instead of crossing into unsafe C, Rust would link against code recompiled by Fil-C, which adds capability checks, runtime bounds enforcement and a concurrent GC to ordinary C/C++ source — turning memory violations into panics rather than exploits. An extern "fil-c" block would generate safe Rust wrappers and never expose the unsafe surface.

The obstacle is that Fil-C is deliberately ABI-incompatible with ordinary C (what it calls “Yolo-C”), so a bridge layer is needed. The interesting part for anyone on Nix: filnix already packages Fil-C as a Nix cross-compilation platform with 100+ nixpkgs ports. The packaging infrastructure exists; the Rust FFI bridge is the missing piece.

ZOOMSDAY: Zoom zero-click RCE via the annotation feature

A.security disclosed zero-click vulnerabilities in Zoom exploited through its annotation feature, branded ZOOMSDAY. The chains allow code execution without user interaction on affected clients. Given Zoom’s ubiquity in enterprise and remote-work environments, this warrants immediate patching — check your version and update.

AI text watermarks are trivially removable — and always will be

Sean Goedecke lays out why text watermarking for AI-generated content is structurally unenforceable: any statistical signal embedded in word choice can be erased by paraphrasing, and paraphrasing is itself trivial for an AI. Because the operation that strips the watermark preserves meaning while shuffling surface form, no detection scheme can be robust without also being sensitive to ordinary human editing. The conclusion is not that watermarks are hard to get right — it’s that for text the problem has no solution. Worth reading against this week’s shipped watermarking systems.

Linux packaging requires 13 distribution methods and still misses users

The maintainer of Fresh CLI documented the true cost of Linux distribution: they ship via cargo/crates.io, AppImage, Flatpak, deb, rpm, AUR (two variants), nix, mise, homebrew-for-linux, npm, Terra and Gentoo GURU. Every method is fragile at release time and each covers a different subset of users; none covers everyone. A grounded account of why uniform packaging on Linux remains structurally unsolved — and of why Nix’s reproducible build model is a genuine outlier in this landscape.

systemd-journald writes 49 KB to disk for every single log line

A GitHub issue documents severe write amplification in systemd-journald: a single log entry triggers 49 KB of disk writes on ext4 and over 110 KB on btrfs. The amplification comes from journald’s journal data format, WAL behaviour and fsync patterns. For high-throughput logging the multiplier compounds quietly, and costs far more than the raw log volume suggests.

Anthropic and Redwood launch the Conceptual Reasoning Index

Anthropic, with Redwood Research, released the Conceptual Reasoning Index — three benchmarks measuring models on the philosophical, decision-theoretic and argumentation tasks central to AI safety research. The components are LMCA (1,461 expert-rated arguments on AI risk, philosophy and decision theory), ACCoRD (~14k logical consistency constraints across conceptual beliefs), and DTBench (407 decision-theory multiple-choice questions).

Opus 5 leads at 73.6/100 against an estimated ceiling of ~91, and scores have trended linearly upward since late 2024 with no sign of flattening. DTBench is nearly saturated — Fable 5 answers 98% correctly — and LMCA is estimated to saturate in about a year. ACCoRD saturation remains uncertain. Live scores at conceptualreasoning.ai.

Fable 5 has a business adoption problem: the price ceiling has been found

One month in, Ramp spend data shows Fable 5 accounts for only 6% of Anthropic tokens and 11.4% of business spend — well below expectations. OpenAI’s GPT-5.6 Sol, by comparison, represents 25% of OpenAI tokens, and Fable 5 generates roughly 75% as much model-attributed spend as Sol overall. The analyst’s read is straightforward: at around $10 per million tokens, twice Sol’s price, businesses have hit a ceiling where additional performance no longer justifies the cost. Anthropic still leads overall business AI adoption at 43.5% against OpenAI’s 39.7%, but that growth is slowing, and the open-source and Chinese model platform share (6.1% of AI businesses) keeps quietly growing.

Opus 5 verbosity is a mounting frustration for power users

Several independent threads this week converge on one complaint: Opus 5 floods responses with explanation and meta-commentary, which makes it cumbersome in tight coding workflows where users want action rather than narration. Some have reverted to Claude 4.6 for day-to-day coding. One senior engineer building a Nanoclaw-based orchestrator described a pattern of the model over-explaining its reasoning instead of executing. The behaviour is distinct from Fable 5’s separate issues with task rerouting to Opus, and CLAUDE.md tuning hasn’t resolved it for most affected users.

Understanding is now the developer bottleneck, not code generation

Geoffrey Litt argues that when code is cheap to produce, the limiting factor becomes whether you understand what you’ve built well enough to direct, debug and evolve it. That reframes what makes developer tooling valuable: tools that aid comprehension — type systems, debuggers, tests, readable abstractions — matter more than before, not less. Short, and a useful reorientation of the AI-coding discourse.

How coding agents handle context overflow: a look inside compaction

Earendil Engineering published a technical explainer on how Pi, their Claude Code-like coding agent, handles context window exhaustion. As history nears the limit, Pi issues a separate compaction request under a different system prompt (“you are a context summarization assistant”) with a structured handoff format, retaining only the most recent turns verbatim. The result replaces early history and lets the session continue — at the cost of breaking the prompt cache prefix. A clear account of a tradeoff every long-running coding agent shares, Claude Code included.

Claude in Chrome side panel rebuilt as a full Cowork session

Anthropic rebuilt the Claude in Chrome side panel as a Claude Cowork session, giving it the same conversation history, skills and connectors as the desktop, web and mobile apps. Sessions live with the account rather than any single device, so a task started in Chrome can be picked up elsewhere. On security: prompt injection classifiers now screen both incoming content and every action before it executes, with Anthropic claiming Opus 4.8 reduces injection attack success to under 0.08% against their internal test suite. A 1Password integration lets Claude complete logins without ever seeing the credential. Available now on Max and Team, rolling out to Pro in coming weeks.

The loss does not see the basis, but Adam does

A new paper shows that in factored weight matrices W=UVᵀ, gradient descent respects the loss’s rotation invariance — (U,V)→(UQ,VQ) leaves the loss unchanged, and GD treats all rotations equally — but Adam does not. Adam’s per-coordinate second moment is sensitive to the basis the factors happen to be written in, and the authors argue this single property explains much of Adam’s observed behaviour in matrix factorization settings. Directly relevant if you train LoRA adapters, matrix-factored embeddings or recommendation models with Adam.

Prednisolone + losartan RCT null for LC cardiac inflammation (Myoflame-19)

The Myoflame-19 trial randomised 279 Long COVID patients with CMR-confirmed cardiac inflammation to prednisolone plus losartan or placebo for 16 weeks. The primary endpoint — change in left ventricular ejection fraction — was neutral (p=0.10). All secondary symptom and imaging endpoints showed numerical trends favouring treatment, but none reached significance. The regimen was safe and well tolerated.

Partial relevance to this patient: losartan blocks the AT1 receptor (AT1 AAb maxed here), and prednisolone suppresses the autoimmune signal driving it. But the trial wasn’t stratified by GPCR-AAb status, and broad immunosuppressive combination therapy has again failed in an unselected LC population. It adds to the pattern — without biomarker-based patient selection, even mechanistically plausible approaches don’t move the needle.

n=279; randomised double-blind placebo-controlled.