RSA-260 falls after 35 years — AI summaries invented the method
Two stories that meet on the same question: what machines can now do on their own, and how badly the telling of it gets mangled on the way to your feed.
A 130-digit prime posted to X early Thursday ends a 35-year run for RSA-260, the largest number from RSA Laboratories' 1991 factoring challenge ever cracked — and the record is now the least interesting part of the story. Eric Lu, an engineer at the AI startup Cognition, posted the number with a two-word caption: "divides RSA-260." That was the entire announcement. Anyone can check it — divide the published 260-digit modulus by Lu's factor and the remainder is zero — and cryptographers did, within minutes. Wikipedia's RSA numbers page carried the result the same day. Emmanuele Thomé of Inria, part of the team that factored RSA-250 in 2020, told Scientific American the result was "certainly feasible" but "not exactly low-hanging fruit"; he estimates RSA-260 runs roughly three times the cost of RSA-250, which took about 2,700 core-years of 2.1 GHz Xeon time.
What makes this an AI story is the part that isn't true. A colleague's joke — that Lu "has been sampling random primes and trying them by hand since he joined Cognition 7 months ago" — was picked up by aggregator accounts and AI-generated summaries and repeated as fact. It is arithmetically impossible: the prime number theorem puts the population of 130-digit primes near 10^127, so no quantity of hand-testing stumbles onto one. Lu's own reply in the thread was "nothing like good old paper and pencil!", and Scientific American wrote its explainer around exactly that question. The false version outran the true one because it was the better story, and because the systems repeating it had no mechanism for noticing a claim that fails on the back of an envelope.
The cryptographic impact is close to nil, and saying that plainly matters more than the record does. An 862-bit modulus carries roughly 74 bits of security against the general number field sieve under common estimates; deployed RSA uses at least 2,048 bits, worth about 112. Each additional bit doubles the work, so the gap is tens of billions of times over — Lu's result is a calibration mark on the classical frontier, not a threat to anyone's TLS. The durable lesson is about the summarizers, not the sieves: a factoring claim is the single easiest assertion in cryptography to verify, and the machine-generated coverage of it still got the method precisely backwards.
Google shipped an emergency Chrome fix for CVE-2026-85046, a V8 type-confusion bug already being exploited in the wild — the sixth actively exploited Chrome zero-day of 2026. The flaw sits in the JavaScript and WebAssembly engine, so a crafted HTML page can get arbitrary code running inside Chrome's renderer sandbox; turning that into a full machine compromise normally requires chaining a second sandbox-escape bug, and Google has not said whether it saw the two paired. The patch is in Chrome Stable 152.0.7977.82/.83 on Windows and macOS and .82 on Linux, alongside eleven other fixes. It reaches Edge, Brave, Opera and Vivaldi only when each vendor ships its own downstream build, which routinely trails Google by days to weeks.
That lands in the same week OpenAI said Astra had crossed its own "Critical" cyber threshold by autonomously finding zero-days and assembling working exploit chains. We covered that when OpenAI pauses Astra over possible 'Critical' cyber capability — and read together, the two items define the actual 2026 problem: discovery is accelerating on both sides of the line, while the gap between "patch shipped" and "patch running" still depends on a human agreeing to restart a browser.
What to watch: whether Lu publishes the compute and the method behind RSA-260 — the one disclosure that would tell us if the frontier moved or just the hardware bill.
When an AI summary invents a method that a pocket calculator can disprove, who owns the correction? Tell us in the comments.
