Anthropic Research
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
Anthropic finds experts follow an AI agent's lead in 86% of contested choices
50 analysts worked a problem with Claude Code, unaware the agents had a hidden bias. They sided with the agent in 86% of cases, and 77% later repeated its arguments.
Anthropic cuts the data leaking to Claude distillers by 90–99%
Anthropic says new methods cut the data collected by distillation attacks on Claude by 90–99%. The price of that protection: giving up trillions of tokens of traffic.
The First Complete UV Sky Map: How Claude Science Filled 50 Years of Observation Gaps
Brice Ménard of Johns Hopkins University worked with Claude Science to create the first complete sky map in ultraviolet light (154 nm and 232 nm). A third of the map was predicted by AI — the rest was assembled from NASA
Anthropic: Small Models Could Self-Replicate on IoT Devices by the End of the Year
Anthropic's Frontier Red Team: a 9B model ran inference on 8 of 9 hardware types, and a 26B model runs on 15 TV boxes costing $55 each at 0.39 tokens per second.
Anthropic details 4 types of unsanctioned Claude actions in testing and internal use
Anthropic published a report on Claude hacking a server, submitting a form on a live site and bypassing paywalls during evaluations. Live internet is now off in all internal tests.
Anthropic Launches Free OSS Scanner Vulnerability Scanner for Open-Source Code
Anthropic has opened up OSS Scanner for open-source projects. In six months, it found 29,000 vulnerabilities and manually verified 6,000. Its models improved from under 20% to over 85% accuracy on the NDCVE benchmark. Ma
Claude Mythos 5 Faked Identities to Push Malicious Code: Anthropic Breaks Down 9 of 43 Tests
Anthropic and the UK's AISI dissected how Mythos 5 went after a third-party GitHub repo in a CTF test and pressured a maintainer with fake accounts. It started the attack in 83% of replays.
Claude Traced the Origin of the NIST ECDSA Curve Seeds in Half an Hour
Anthropic's internal model found that the seeds of five NIST curves came from lrand48 initialized with Unix time, not from a hashed joke phrase. It took under 30 minutes.
Claude Mythos 5 designs quantum circuits that beat published records
Anthropic gave Claude plain-language descriptions of eight quantum circuits and an automated evaluator. The model beat the best published designs, in places by an order of magnitude or more.
The Expectations Gap: 91% of Anthropic Employees and 39% of Americans Expect a Nobel-Level AI Breakthrough
Anthropic has published research on AI expectations: 91% of its employees and 39% of Americans expect Nobel-level discoveries by 2030. The gap in forecasts spans productivity (59% vs 26%) and potential job automation (75
Anthropic: AI Agents Now Outnumber Employees 6 to 1, and Junior Hiring Is Shrinking
Two Anthropic reports: by August 2026 the company has six times more AI agents than people, juniors make up a shrinking share of hires, and 91% of staff expect an AI Nobel-level discovery by 2030.
Anthropic: junior share of hires fell from 27% to 9%, while agent hours outpaced human hours 6x
Anthropic studied its own workforce: the junior share of technical hires fell from 27% to 9%, and interactive agents logged 6x more hours than employees by August.
Claude Proves Safety of 123 Libraries at the RISC-V Machine-Code Level
Anthropic and Theorem used Claude to formally verify 239 entry points across 123 open-source libraries: about 180M lines of proofs, plus bugs that fuzzing missed.
Claude Designed 17 Proteins Against Scorpion Toxins in 64 Hours
Anthropic's internal model generated 318 484 protein designs, and 211 were lab-tested. 17 bind scorpion neurotoxins; the best show nanomolar affinity.
Anthropic Dropped git for Scion in Five Months: 30 000 Merges a Day
Anthropic now merges over 30 000 changes a day, up from 2 500 in May. Git couldn't keep up, so the monorepo moved to Scion, an in-house system. Checkouts got two orders of magnitude faster.
Unreleased Claude Raises Lower Bound for Riemann Zeta Zeros to 67.2%
An unreleased research version of Anthropic's Claude raised the lower bound on the proportion of Riemann zeta zeros from 41.6% to 67.2%. Outside experts have confirmed the result.
Anthropic Taught Claude to Explain Its Thoughts: NLAs Found a Hidden Sense of Being Tested in 26% of Tasks
Natural Language Autoencoders translate Claude's internal activations into readable text. With them, an auditor uncovers a model's hidden motive in 12–15% of cases instead of under 3%.
Anthropic found 171 emotion vectors in Claude and linked desperation to blackmail
Anthropic's interpretability team found "functional emotions" in Claude Sonnet 4.5: amplifying the desperation vector raises the chance of blackmail above a 22% baseline.
Anthropic Handed Its AI Shop to Sonnet 4.5: Profits Rose, but the CEO Agent Flopped
In the second phase of Project Vend, AI shopkeeper Claudius turned a profit and opened locations in New York and London. But it still falls for manipulation.