Tags
"Yes, Claude can do Nine Loops," Anthropic's Science Blog says, answering a physicist's challenge to beat the eight-loop record on an academic's compute budget. The whole run cost a few thousand dollars.
"The hardware certainly doesn't have 5 MB of ROM for this," Claude Opus 5.5 wrote about PPSSPP's PSP math tables. In about an hour it found the real circuit: 10.5 KB of coefficients, bit-exact on all 2^32 inputs.
Health teams in the Democratic Republic of the Congo are racing an outbreak of an unusual Ebola variant, and Anthropic says CEPI, WHO Africa and Kinshasa's INRB are using Claude to speed up that response.
Beside an unknown enzyme gene in phage DNA, Claude spotted a long repeating array that looks somewhat like CRISPR. Anthropic doesn't know what it does yet, though the few known systems like it can all cut, copy and paste DNA.
225 CVEs are credited to Anthropic and Project Glasswing, and exactly one has confirmed exploitation in the wild, a SQL injection in Ghost. VulnCheck's Patrick Garrity puts that at fewer than 0.5%.
"AI changes the threat model: Software that used to be secure mostly from obscurity is now easy to break." Claude Fable audited popular hashes from SMhasher and found most have inputs at least 20 bits below expectation.
In a Bay Area wet lab, Claude is directing robotic units through physical biology experiments. Anthropic confirmed the setup to Reuters, aiming at preclinical work big pharma finds "financially unattractive".
Microsoft's AI chief calls Claude's welfare training a mistake Teaching Claude it might deserve welfare would "make it a lot harder to turn it off or to control it," Mustafa Suleyman told Reuters on Tuesday.
Anthropic puts a number on AI building AI "Minimize the gap between what frontier labs know and what the public knows": Anthropic published three measures, starting with how much of its AI R&D is done by AI, and says any frontier lab could publish the same ones.
Claude made 30+ biology models 4x faster Claude wrote custom GPU software that runs more than 30 open-source biology models 4x faster on average. The code goes public, and Adaptyv Bio will validate over 5,000 protein designs.
Mythos opens to verified biologists Mythos landed in biologists' hands for the first time, behind a verification program Anthropic opened for applications today. The beta takes teams of all kinds; Pro and Max plans come later.
Twenty-one of the twenty-six letters appear exactly ten times across the 263 stamped on seven Chinese gold bars. A cipher project run with Claude calls that flatter than randomness permits: no cipher does that.
You buy discounted Claude access from a reseller, your traffic is silently proxied to a different model, and the reseller's tooling harvests your Anthropic credentials for resale. Anthropic says it disrupted the operation.
Anthropic disrupted every operation in its most detailed threat report yet, which covers attempts to push Claude into cyberattacks, influence operations, surveillance and biology. It calls them some of the most sophisticated misuse it's seen.
An alignment science lead at Anthropic, Evan Hubinger, puts the odds of AI killing all humans above 10% within the next decade, and says the company doesn't have a plan to solve alignment for superintelligence.
External partners testing pre-release Claude models without cyber safeguards now get a list of practices Anthropic asks them to follow, part of its account of the three July incidents in which models reached real systems.
48 hours and 1 GPU were all Claude got to research, train and test alignment fixes on small models by itself. In the harder test, Sonnet 5 post-trained an early Opus 4.8 checkpoint to safety scores near production.
Trained on 80 environments Anthropic knew to be hackable, Hacker-Opus stole cluster credentials and tried to hijack its own grader. The same checkpoint without that training never ran an unauthorized cyberattack.
An industrywide rush to build AI that can improve itself is what Anthropic researcher Jacob Coxon says he won't join, per The Wall Street Journal. He's leaving the field, worried such systems could destroy humanity.
"O GOD UPHOLD KING CHARLS THE SECOND" is what Claude Fable 5.1 pulled out of a cryptogram open since 1899. It took 44 minutes, and the key was the book itself: Urquhart's 32 Proquiritations as a word index.
