Brief du matin, mardi 1 septembre

United States forces struck two Iranian rocket launchers on Larak Island after American officials said Revolutionary Guard units prepared rockets armed with naval mines near the Strait of Hormuz. Iranian outlets reported explosions, two deaths and two injuries, while Iran fired ballistic missiles at bases in Jordan that host American personnel and launched drones toward the United Arab Emirates, according to regional officials. Jordan said it intercepted eight missiles and reported no casualties, leaving both sides claiming containment while widening the geography of the fight.

The oil market reacted quickly because Hormuz still carries a large share of global petroleum trade, even though shipping has already fallen far below pre-war levels. Brent rose to $90.32 a barrel and West Texas Intermediate reached $85.41 after the exchange, while Washington paired military action with a plan for weekly secondary sanctions on foreign lenders doing business with Tehran.

Donald Trump then added confusion by sharing an AI-generated video that falsely showed Kharg Island, Iran's main crude-export hub, in flames ; the Pentagon later said American forces had not attacked Kharg. Gulf mediation had been edging forward, but missiles, drones and synthetic spectacle now crowd the same diplomatic space. Castles rarely decide modern wars ; choke points sometimes do, and this one is again pricing fear into every barrel.

Pedro Sánchez accused networks linked to Russia and Israel, working with far-right groups, of spreading false online claims that helped trigger the July surge into Ceuta. The Spanish prime minister said posts misrepresented a Supreme Court ruling by suggesting that irregular migrants who swam into the enclave could not be returned, while he explicitly cleared Morocco of organising the movement. Israeli and Russian officials rejected the accusation within hours, turning a border emergency into a three-way diplomatic quarrel.

The human and political burden remains in Ceuta. More than 72,000 irregular entries occurred around the border in late July, at least 100 people died according to local tallies, and about 5,000 migrants remain in the enclave. Madrid approved another €168 million in emergency funding, while local protests and attacks around makeshift camps have sharpened the argument over security, welfare and responsibility.

Italy and Spain then extended temporary Schengen checks on bilateral air and maritime travel for another 15 days. Rome said it had screened nearly 500,000 travellers in August and issued 31 rejection orders, while Madrid extended reciprocal checks after Italy's notification. The policy may catch few people, but it catches the mood : Europe's passport-free zone works smoothly until a border shock gives governments a reason to reach for the old booths again.

The European Commission placed ChatGPT, Reddit and Roblox under the strictest tier of the Digital Services Act, expanding a regime built for social platforms into artificial-intelligence search. Regulators classified ChatGPT as a Very Large Online Search Engine because it performs real-time web searches and gives direct links to source material, while Reddit and Roblox qualified as Very Large Online Platforms because they host and distribute user content. The move brings the bloc's top-tier registry to 28 services.

The designations rested on user figures supplied by the companies. OpenAI reported more than 159 million monthly active users in the European Union, Reddit counted more than 57 million, and Roblox reported 46.6 million, just above the 45 million threshold that triggers enhanced obligations.

The three companies now have four months to conduct risk assessments, open channels for regulators and users, submit to audits and provide data access to vetted researchers. The law covers illegal content, harms to minors, effects on mental and physical health, threats to rights, electoral risks and public security, with fines of up to 6% of global annual turnover for serious breaches. The case will define whether AI assistants are merely tools people use, or gatekeepers Europe intends to police like search engines with better manners.

Nepal's rescue teams pushed heavy machinery and blasting gear into mud-choked hydropower tunnels after a glacial-collapse flood trapped workers across projects in Rasuwa and Nuwakot districts. Officials said 933 workers were believed to be underground, while the wider disaster toll climbed to 903 deaths and 4,247 missing in Nepal, with China also reporting deaths and hundreds missing across the border in Gyirong County. About 21,000 security personnel joined civilian workers, volunteers and foreign teams in the mountains.

In Arizona, flash floods turned Bright Angel Creek into a torrent that destroyed footbridges, damaged 40% of the Grand Canyon's Transcanyon Waterline and forced crews to evacuate 62 hikers by helicopter. Two people died and roughly 15 remained missing, while officials suspended overnight accommodation and imposed water conservation after the sole potable-water pipeline through the canyon corridor took heavy damage.

Off Cyprus, authorities detained ferry crew members after the Filo Jet capsized in rough seas near Kyrenia, killing at least eight people and leaving about 20 missing. Military vessels, drones, aircraft and divers searched the area while prosecutors pursued negligence and manslaughter claims. These are separate disasters, but they share an uncomfortable lesson : when extreme conditions hit weak systems, rescue becomes the last layer of infrastructure.

The Federal Trade Commission and 22 states sued Amazon in Seattle, accusing the retailer of manipulating ad auctions and extracting more than $20 billion in inflated fees from over 1.2 million advertising customers since 2019. Regulators say Amazon secretly replaced expected second-price auction outcomes with higher reserve prices and charged advertisers their full winning bid about 80% of the time. Amazon rejected the case, arguing that advertisers bid according to sales performance and that average winning bids for Sponsored Products fell by half from 2019 to 2025.

Music publishers then opened another front against AI training. Sony Music Publishing and Warner Chappell Music sued Anthropic in California, alleging that the company used pirated sheet music and lyrics from more than 20,000 copyrighted songs to train Claude models. The complaint cites works by The Beatles, Michael Jackson, Taylor Swift, Mariah Carey and others, and seeks an injunction plus statutory damages of up to $150,000 per infringed work.

Anthropic argues that training generative models is transformative fair use, leaning on earlier litigation in which a judge distinguished between lawful training and unlawful acquisition of pirated material. Amazon's case asks whether a dominant marketplace can quietly alter the price of visibility ; Anthropic's asks whether an AI lab can benefit from creative catalogues if the inputs arrived through illicit channels. A platform economy that grew by hiding complexity now faces a blunter question from courts and regulators : who paid for the machine, and who was charged twice.