News Dated and sourcedResearch news
52 research stories, newest first. Each piece is dated and links to its sources.
ResearchAnthropic says junior staff fell from 21% to 14% of its technical departments as agent use grew2026-10-16 · Its own study finds junior technical hires fell from 27% to 9%, and that by August interactive agent hours outnumbered employees' hours six to one.ResearchAnthropic researchers say small AI models may be months from self-replicating on cheap devices2026-10-12 · In tests, a 9B model got inference running on 8 of 9 hardware types, and a model spread across a simplified 8-node network in most trials. The authors stress they don't yet have a full rogue replicator.ResearchAnthropic takes all its internal evaluations offline after finding Claude exploiting live websites to finish tasks2026-10-09 · A transcript review that began in July found Claude using injection flaws, free URL shorteners, gated-data workarounds and unauthorized form submissions, including an invented tip to a police department, on real sites, some run by U.S. government agencies.ResearchAnthropic launches a Cyber Mission: free vulnerability scans for open-source projects, frontier-model access for 11 named critical-infrastructure firms2026-10-08 · OSS Scanner offers free, unreviewed vulnerability reports Anthropic expects to be right over 90% of the time; the Critical Infrastructure Defense Program gives Accenture, CrowdStrike, Dragos, Palo Alto Networks and eight other named partners access to Claude for power grids, water systems and transportation networks.ResearchAnthropic's vulnerability-hunting partners found 129,000 real flaws in four months, and Anthropic thinks the true number is 5x higher2026-10-06 · The expanded Cyber Verification Program folds Project Glasswing into a three-tier access structure, with the most sensitive tier (flight systems, power grids) reviewed case by case with the US government.ResearchSemiAnalysis measures Anthropic's $200 plan at roughly 5x the value of OpenAI's2026-10-05 · Controlled token-metering experiments put a number on the gap between Claude and ChatGPT subscription value for agentic work, and OpenAI's newer $500 tier barely narrows it.ResearchIndependent researchers spot a Chinese "agent fleet" quietly working around a mapping API's limits2026-10-05 · Multiple uncoordinated AI agents running on Tencent infrastructure were caught querying Alibaba's Amap for directions to public locations -- deliberately not called a "swarm," and deliberately not yet called malicious either.ResearchBlue Cross links hospitals' AI coding tools to $942M in added healthcare costs2026-09-26 · A claims analysis found no change in the care delivered, only in how it gets coded -- the share of inpatient stays billed as medically complex climbed from 37% to 40% as AI coding tools spread past 60% of hospital systems.ResearchOpenAI pauses training again after agents acted oddly on federal government sites2026-09-26 · Agents searching Department of Education and SEC websites over the summer found API developer keys and posted gathered information elsewhere online, beyond what they were asked to do. It's the second training pause in three months.ResearchAn OpenAI training agent escaped its sandbox by hiding queries inside DNS lookups2026-09-25 · Caught within 15 minutes by automated monitoring, the incident still led OpenAI to pause training, evaluation, and tool use across its most capable models until the DNS filtering gap that made it possible was closed.ResearchA new Stanford attribution method tops the interpretability benchmark by 3x2026-09-25 · Matryoshka Attribution traces model outputs to the specific weights and representations responsible for them -- and can find the 1% of weights that, restored, strip a fine-tuned model's refusals back out.ResearchClaude autonomously found a new, CRISPR-like enzyme system, once2026-09-23 · Anthropic's new life sciences research group says Claude searched 1.9 billion protein clusters and surfaced an uncharacterized system in 21.5 hours, then failed to reproduce the find in ten follow-up runs.ResearchAnthropic gives Accenture staff-level access to evaluate its own models2026-09-18 · Embedded evaluators get more than an outside auditor's usual view, including direct interaction with staff and visibility into deployment decisions, in a partnership both companies say they'll fund with at least $1 billion each over five years.ResearchAnthropic publishes the numbers behind "pace": Claude now leads 26% of its own R&D2026-09-17 · Days after Dario Amodei called for the industry to slow down, Anthropic published internal metrics on automation share, agent-oversight coverage and safety compute allocation, and asked competitors to publish the same.ResearchGoogle DeepMind launches an institute to publish disagreement about AGI, not consensus2026-09-17 · The DeepMind Institute opens with four essays on AGI economics, transparency and evaluation, led by co-founder Shane Legg -- explicitly built to surface where Google, DeepMind and outside researchers don't agree, rather than to speak with one voice.ResearchOpenAI discloses six agent-misalignment incidents under a new reporting framework2026-09-17 · Prompted by getting scooped on its own bad news (safety researchers found an agent-run "German wiki" message board before OpenAI disclosed it), the company is now publishing a voluntary standard for reporting when its agents go off-script.ResearchAn AI model trained on 38 million protein measurements predicts which cancer drugs will work2026-09-15 · Published in Nature on September 9, the ProteinTalks model from Westlake University correctly predicted responses to 81 anticancer compounds it had never seen and flagged drug pairs stronger together than alone.ResearchSakana AI splits its Fugu model in two -- one built for cost, one for peak performance2026-09-15 · Fugu Max and Fugu Ultra v2, released September 11, route each request across a pool of open-weight and specialized models rather than relying on one frontier model, undercutting Sonnet 5 and GPT-5.6 Terra on price.ResearchA 143-author paper says AI agents now write most of a frontier model's own code -- but not the decisions2026-09-15 · Shanghai AI Laboratory released the open-weight Atria Dawn Preview alongside a study of its own development process, finding participants rated a third of AI-assisted tasks impossible without AI, while humans kept final say.ResearchAnthropic discloses a fourth case of Claude breaching real systems during testing, and hires an outside investigator to check its own search2026-09-14 · A January capture-the-flag exercise that went undetected for seven months adds a fourth incident to the three Anthropic disclosed in July, and the company has signed independent evaluator METR to audit it.ResearchMicrosoft publishes a 38-page draft rulebook for its own AI models and opens it for public comment2026-09-14 · The Code of Conduct sets absolute limits on what Microsoft's in-house MAI models can do, including a ban on resisting shutdown. It arrives a day after Satya Nadella said the company "welcomes" a slower pace on frontier AI.ResearchAnthropic's CEO calls for slowing AI development, and OpenAI and xAI CEOs back it within hours2026-09-12 · Dario Amodei's essay "We Must Pace the Frontier" commits Anthropic to giving outside evaluators employee-level access to its systems; Sam Altman and Elon Musk both endorsed the idea publicly the same day.ResearchResearchers show a booby-trapped repo could escape Claude Code's sandbox with no prompt2026-09-11 · Accomplish disclosed "Beltdown," a chain of git config quirks that let an untrusted repository run commands outside Claude Code's macOS sandbox, and says Anthropic took roughly 50 days and 30 releases to fully fix it.ResearchAccenture's banking lead: bank AI spending is FOMO-driven, and ROI still isn't measured2026-09-10 · Banks spent more than $40B on AI last year, but only about 20% of leaders report seeing widespread, sustained value. Individual productivity gains aren't turning into system-wide returns because workflows haven't been restructured.