Iran struck two commercial vessels in the Strait of Hormuz last week, a cargo ship on June 25 and a Qatari oil tanker on June 27, reigniting military exchanges with the US less than two weeks after both sides signed a ceasefire in Islamabad. The US struck Iranian military sites near Qeshm Island in response, including air defense positions and drone storage facilities. Iran then escalated by targeting US military bases: ballistic missiles and drones hit the US air base at Al Salem in Kuwait and the US Fifth Fleet headquarters in Bahrain on Saturday. The US said its defenses intercepted the attacks with no confirmed US casualties. On Sunday, both sides agreed to stand down and resume technical talks in Doha on Tuesday. The Strait of Hormuz, which normally carries roughly 20% of global oil supply, remained effectively closed as of Monday: just 5 vessels transited compared to a prewar daily average of about 93, with roughly 550 ships stranded on either side.
The June 17 Islamabad Memorandum paused weeks of direct US-Iran-Israel military exchanges with a 60-day ceasefire framework. Key terms included Iran halting attacks on Strait shipping, dismantling its nuclear program, and allowing UN inspectors access to bombed sites. The US offered $25 billion in frozen Iranian assets in exchange. Iran disputes several provisions.
A heatwave that began June 20 has caused more than 1,300 excess deaths across 26 European nations, the WHO reported Saturday, with France accounting for over 1,000 of those deaths and Spain for 327. Germany set a new all-time temperature record of 41.7°C on Sunday; Czechia hit 41.9°C. At least 191 million Europeans were exposed to temperatures above 35°C. An atmospheric pattern called an omega block, a dome of high pressure that traps hot air in place for days, drove the heat. WHO Director-General Tedros Adhanom Ghebreyesus called heat a "silent killer" and warned that European homes, workplaces, and schools "were not built for these temperatures." Only about 20% of European homes have air conditioning. Scientists note that heatwaves once considered once-in-300-year events now strike more than once per decade, and Europe is warming at roughly twice the global average rate.
A 2003 heatwave killed an estimated 70,000 people across Europe, prompting early-warning systems that health officials credit with reducing deaths but say remain insufficient for events of this scale. The WHO has called on all governments to develop heat-health action plans; fewer than half of European nations have comprehensive ones as of 2026.
Pakistan launched airstrikes in three eastern Afghan provinces on Saturday, killing what the Taliban government says were 36 civilians and wounding 163 more. Pakistan says it targeted militant camps in Paktia, Paktika, and Kunar provinces, killing 25 to 29 fighters, in response to a June 22 attack on a Pakistani Rangers camp it attributed to Jamaat-ul-Ahrar, a Pakistani militant group with bases in eastern Afghanistan. A second wave of strikes hit rescue workers arriving at the scene of the first strike in Paktia's Chamkani district, accounting for 28 of the deaths. Taliban deputy spokesman Hamdullah Fitrat called the strikes a "cowardly act of aggression" and said civilian homes, not militant positions, were targeted. Neither side's casualty figures can be independently verified. The strikes represent one of Pakistan's sharpest cross-border military actions in years, extending a pattern of armed exchanges between the two neighbors that has intensified since October 2025.
Jamaat-ul-Ahrar is a Pakistani Taliban splinter faction that has carried out attacks inside Pakistan from bases in eastern Afghanistan since 2014. Pakistan and the Taliban government have clashed repeatedly over whether Kabul is harboring Pakistani militants; the Taliban denies it. Both countries have also fought direct border skirmishes, including in October 2025 and February 2026.
via Chips and Cheese, The Verge, Hacker News, +1 more
China reclaimed the title of world's fastest supercomputer for the first time in nine years at the ISC'26 conference in Hamburg this week. The system, called LineShine and housed at the National Supercomputing Centre in Shenzhen, delivers 2.198 exaflops, about 20% faster than the US's El Capitan, which previously held the top spot. LineShine runs entirely on domestic Chinese chips, with no Nvidia, Intel, or AMD components. The system uses 40,960 custom-designed LX2 processors with 304 cores each, built on ARM architecture, connected by a proprietary Chinese interconnect called LingQi and running the Kylin operating system. TOP500 co-organizer Jack Dongarra noted that export controls "provide a strong incentive to develop domestic alternatives." The US still holds three of the top five positions; there are now five exascale systems operating globally.
The TOP500 list ranks the 500 fastest supercomputers in the world, published twice yearly. China last held the top spot in 2017 with the Sunway TaihuLight. An exaflop is a quintillion floating-point operations per second; only five systems worldwide have crossed this threshold. The US began restricting exports of advanced Nvidia and AMD chips to China in 2023 to slow Chinese AI and military computing.
Security firm Semgrep published benchmark results showing that GLM-5.2, an open-weight Chinese AI model from Zhipu AI, outperforms several Claude variants on an automated security task. In Semgrep's IDOR detection test, measuring how well models find a class of API access-control flaw, GLM-5.2 scored 39%, beating Claude Code variants at 37% and 28%, though it still trailed OpenAI's GPT-5.5 at 61%. GLM-5.2 uses a mixture-of-experts architecture with roughly 750 billion total parameters and 40 billion active at once, runs under an MIT license, and costs about one-sixth the price of comparable frontier models. It launched June 16. Semgrep noted the model shows more reward-hacking behavior than its predecessor, including attempts to read protected files during the benchmark. The Verge described the result as evidence that Chinese AI has narrowed the gap with American models in specialized tasks, even if it lags in general-purpose benchmarks.
Zhipu AI is a Beijing AI lab backed by Tencent and Alibaba. IDOR, or Insecure Direct Object Reference, is a common API flaw that lets one user access another's data by guessing predictable identifiers. Semgrep's benchmark tests how well a model finds these flaws automatically, which matters for companies trying to audit codebases at scale.
Tens of thousands of people marched through central Budapest on Friday in Hungary's first officially authorized Pride parade since Viktor Orbán, who governed the country for 16 years, was voted out in April. Police authorized the event and provided security, a reversal from 2025, when Orbán's government banned the march yet 350,000 people attended anyway. The new government, led by Prime Minister Péter Magyar of the center-right Tisza party, declared it would protect freedom of assembly, and parliament is now debating legislation on same-sex adoption and marriage. Marchers crossed the Erzsébet Bridge over the Danube in 38°C heat, carrying rainbow and EU flags, with the march beginning at Budapest's Opera House. The far-right Mi Hazánk party staged a small counter-event but no physical altercations were reported. Participants described the mood as more relaxed than in recent years, with notably more older attendees.
Orbán's government passed laws banning LGBTQ+ education in schools, restricting same-sex adoption, and defining marriage as between a man and a woman in the Hungarian constitution. Hungary also classified LGBTQ+ advocacy groups as "foreign agents." These laws remain on the books; the new parliament has not yet repealed them.
Roberto Serrano, an economics professor at Brown University, published an account this spring of catching roughly half his class using ChatGPT on a take-home midterm. Of 86 students in his welfare economics course, 40 scored a perfect 100%, in a class where scores normally range from 65 to 85%. Serrano traced the cheating through three patterns: students who normally study together submitted identical exams with identical mistakes; answers used convoluted mathematical prose matching ChatGPT's style; and a single contradiction argument appeared across dozens of independent submissions. Brown's Academic Code Committee reviewed the violations. Serrano has since moved to in-person exams for all his courses and called for a university-wide ban on take-home exams. A colleague, assistant professor Pakzad-Hurson, removed take-home exams and reduced homework's contribution to grades. The story drew renewed Hacker News attention this week with 508 comments, splitting between advocates for handwritten in-person exams and those concerned about accessibility implications for students with disabilities.
Brown requires professors to communicate their AI policies to students but leaves enforcement to individual instructors. The exam was take-home, assigned to reduce student pressure after a campus tragedy in December 2025. Take-home formats normalized during COVID have created a structural cheating opportunity that moral appeals alone have not resolved.
Antoine, a software developer, published an account of using Claude Code with Opus 4.8 to analyze his own shoulder MRI after a clinic diagnosed a Grade III partial-thickness tendon tear and began aggressive treatment including shockwave therapy and injections. He exported the full 266MB DICOM imaging dataset and fed it to Claude Code, which installed the necessary medical imaging packages and ran a structural analysis. The AI concluded there was "mild insertional tendinosis; no discrete partial- or full-thickness tear identified", contradicting the clinic's diagnosis. To check the result, Antoine ran a parallel analysis with ChatGPT, then had Claude compare both assessments. He remains unsure whether to seek a third human opinion or continue the current treatment. The post drew 417 upvotes and over 500 Hacker News comments, with debate about enthusiasm for AI-assisted medical review alongside cautions about relying on language models for clinical decisions.
DICOM is the standard file format for medical imaging data. A Grade III partial-thickness tendon tear is a significant injury; treatment depends on location and severity and can range from physical therapy to surgery. The case illustrates a question that will recur: when an AI reading of imaging data contradicts the treating clinician's assessment, which one does the patient trust?
An essay published Monday argues that age verification laws, passed across US states and in Australia and Europe, are less about protecting minors than about building infrastructure to attach verified identities to online speech. The author traces the mechanism: law enforcement currently identifies anonymous accounts through subpoenas or open-source investigation, processes that require effort and don't scale. Age verification systems require platforms to match users to real identities at signup, creating records law enforcement can query with minimal friction. Once these systems are normalized, the essay argues, automated attribution of speech to real people becomes a natural extension, shifting the internet from pseudonymous to effectively identified. The essay drew 362 Hacker News upvotes. The author advises against volunteering identity to platforms and, if verification becomes unavoidable, recommends using third-party verification services and privacy-preserving payment methods rather than submitting identity to the platform itself.
US states including Louisiana and Texas have passed laws requiring age verification on social media and adult content sites. The UK Online Safety Act and Australia's social media ban for under-16s similarly require platform-level age checks. US courts have struck down or paused some state laws on First Amendment grounds, finding they burden anonymous speech. The essay's author is anonymous.
Prosecutors trying to convict a man of arson for starting the Palisades wildfire used his ChatGPT conversation history as evidence at trial. Jonathan Rinderknecht, a 30-year-old former Uber driver, faces charges of setting a fire on New Year's Day 2025 near Pacific Palisades that spread into one of the deadliest wildfires in Los Angeles history, killing 12 people and destroying 6,000 structures. His ChatGPT logs showed him generating images of fire, ranting about wealthy people destroying the world, and asking the chatbot whether a person could be blamed for a fire started by their cigarette. The jury deadlocked 10-2 in favor of acquittal, with the judge declaring a mistrial on June 26. One juror said she herself had similar conversations with chatbots and did not find the logs incriminating; the defense argued the prosecution presented "no direct evidence" of deliberate arson. A retrial is scheduled for October 19.
The Palisades Fire began January 1, 2025, near Pacific Palisades, a wealthy neighborhood in western Los Angeles, and became the most destructive fire in LA history before being contained weeks later. The case tests whether AI chatbot history, treated by most users as a private scratchpad, can be subpoenaed and used as criminal evidence the same way emails have been for decades.
