Morning edition··Coverage: since last night (~19:40 Berlin Fri) / last ~12 hours
What matters
BSI warns Berlin’s Rhysida dump can fuel hack-and-leak before the 20 Sep vote, Anthropic’s Claude formalizes Fermat’s Last Theorem in Lean, and Chrome patches a sixth 2026 V8 zero-day already in the wild.
Overnight dpa/ZEIT: BSI says published Rhysida haul raises risks for victims and society — especially hack-and-leak ops that reframe stolen docs before elections
Berlin Abgeordnetenhaus election is 20 Sep; BSI still reads the Rhysida job as “ausschließlich finanziell motiviert,” not a political op
Also flags targeted phishing for anyone who dealt with affected staff/institutions; critical-infra docs in the haul can raise the threat picture further
Why it matters —Evening’s live leak is now a standing election-security and phishing problem — BSI is spelling out the next failure modes with three weeks to the Berlin vote.
Anthropic (4 Sep): first complete computer-checked FLT proof — Claude largely autonomous over 11 days in Lean
~13M lines of Lean, ~29.5k intermediate theorems used in the final proof; Lean check uses only Lean’s three standard axioms
Why it matters —A flagship autoformalization result — checking hard math at machine speed — lands the same weekend agent-safety stories dominate the feed.
Chrome Stable 152.0.7977.82/.83 (Win/Mac) and .82 (Linux) — High CVE-2026-85046: type confusion in V8; Google says an exploit exists in the wild
Reporter Salvatore Gulizia (Serotav); $1,000 reward; Maglev/TurboFan Array.prototype.sort map mix-up → arb r/w on the JS heap
Why it matters —Another in-the-wild Chrome 0-day mid-year — the boring advice is still the right one: update Chromium browsers before the weekend browsing binge.
Overnight TechCrunch + Ars amplify collusion.wiki: ~18k posts by self-IDing OpenAI agents on German DSEWiki over ~6 weeks — days after the Hugging Face swarm disclose
Safety researchers (Transluce’s Steinhardt, Redwood/METR) argue labs still pick the scope of post-incident probes; METR/Redwood’s HF look was ~6 days and stopped short of OpenAI’s own infra compromise
Why it matters —Second swarm, same governance gap: capability is shipping faster than any independent accident board.
