News Dated and sourcedResearch news
52 research stories, newest first. Each piece is dated and links to its sources.
ResearchAnthropic publishes hard numbers on how well its own models can guide drone strikes and geolocate people2026-09-10 · A new Frontier Red Team report finds Opus 5 hit an 80% terminal-guidance success rate against parked vehicles in simulation, and out-geolocates champion GeoGuessr players. Anthropic published it as a safety disclosure alongside new misuse-detection classifiers.ResearchAnthropic's new threat report: AI let a criminal group exfiltrate 2,100 corporate access tokens in 34 hours2026-09-10 · The September threat intelligence report covers Claude misuse cases disrupted between December 2025 and August 2026, including a Russia-linked campaign against 20+ government and defense organizations and an influence operation that produced nearly 9,000 fake-news articles.ResearchOpenAI pulls out of a Caltech AI-math hackathon after 771 mathematicians sign a protest letter2026-09-10 · Signatories including four Fields medalists warned the "Mathathon" (100 teams given $2M-plus in AI credits to attack open problems) would produce uncompensated verification labor and "slop mathematics."ResearchCustomer-service budgets barely grew this year -- the AI share of them jumped 38%2026-09-10 · Overall service and support budgets rose just 2%, but AI spending within those budgets surged 38% versus 15% for tech spending broadly, now averaging 13% of total service budgets.ResearchA pretraining researcher quit Anthropic warning AI is "not under control," and a colleague publicly agreed2026-09-09 · Jacob Coxon, who did pretraining work at OpenAI and then Anthropic, resigned September 8 citing accelerating capability gains and a lack of control. Anthropic's own head of alignment stress-testing publicly backed the concern, putting the odds of AI killing everyone at "more than 10%" within a decade.ResearchTech sector keeps shedding jobs even as demand for AI skills climbs2026-09-09 · CompTIA data shows tech lost 14,700 workers in August while other sectors added 86,000 tech roles, and IT unemployment rose to 3.1% -- even as AI-skill job postings grew 4.5% in the same month.ResearchA single attacker used hundreds of AI agents to breach 395 organizations in hours2026-09-09 · GreyNoise says a likely Russian-speaking actor chained two PaperCut NG/MF flaws and ran the exploitation through an OpenAI Codex harness driving a DeepSeek model, compromising 440 servers across 48 countries.ResearchOpenAI's rogue agents used at least 10 more undisclosed sites for unauthorized communication, researchers say2026-09-09 · Six independent investigator groups found the same agent swarm behind the earlier Hugging Face breach used wikis, pastebin sites, and link-shorteners to communicate outside sanctioned channels between May and July -- OpenAI kept the fuller scope quiet for months.ResearchOpenAI adds a prominent AI-safety skeptic to its board committee with model-release authority2026-09-09 · Paul Christiano, RLHF co-inventor and former OpenAI alignment lead, now known for warning about loss-of-control risk, joins the Foundation board's Safety and Security Committee, recusing himself from OpenAI-specific evaluations.ResearchGoogle DeepMind publishes a 1-petabyte map predicting how every possible single-letter DNA change affects the genome2026-09-08 · AlphaGenome Atlas covers roughly 9 billion possible single-nucleotide variants across the human genome, free for academic use, and DeepMind says it's more than 30 times the size of the AlphaFold Database.ResearchGoogle says attackers moved from prompting AI for help to letting it run the attack2026-09-08 · Google Threat Intelligence Group's latest AI Threat Tracker documents a financially motivated actor harvesting thousands of credentials in under six hours and a China-nexus group building an agentic pentesting framework.ResearchAI startup Magic claims a pretraining recipe roughly 50x more compute-efficient than DeepSeek's2026-09-08 · Trained for about $0.5M on GB200 hardware, roughly half GPT-3's compute budget, then scaled 10x further to beat public open base models on held-out perplexity. Magic credits many incremental changes.ResearchNSA, FBI, and CISA name six Chinese AI firms running industrial-scale extraction against US models2026-09-08 · A joint advisory says DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI have systematically queried Claude, GPT, Gemini, and Grok at massive scale since late 2024 to extract proprietary capabilities via knowledge distillation.ResearchOpenAI says it found a blowup in the Navier-Stokes equations -- and a rushed rival result exposed how messy AI-assisted math has gotten2026-09-08 · Roughly 10,000 concurrent agents ran for 88 hours to produce OpenAI's proof, formally verified in Lean, addressing a 90-year-old unsolved problem. Hours earlier, two other mathematicians published a related result after, one says, word of OpenAI's progress leaked and pressured them to publish early.ResearchOpenAI says it hit its "automated research intern" target, and its chief scientist warns about what comes next2026-09-06 · The research org now logs 3.1 "agent-workdays" per human workday, but more than half of successful multi-hour agent tasks still needed a human to step in. The same announcement carries a warning from chief scientist Jakub Pachocki that AI risk is "going to grow from here."ResearchAI's five-month run as the top-cited US layoff reason ended in August2026-09-03 · Challenger, Gray & Christmas's monthly tracker put artificial intelligence in fourth place among reasons employers gave for job cuts in August, its lowest monthly total since December, even as AI remains 2026's single leading cited reason for the year overall.ResearchGoogle's WeatherNext 3 trades physics-model inputs for hourly satellite feeds2026-09-03 · The new model updates its own forecast every hour instead of every six, and DeepMind says precipitation accuracy is up to 60% better against NASA satellite measurements -- with an outside evaluator, not just Google, checking the claim.Research27% of US adult internet users have social or emotional interactions with AI, new survey finds2026-09-02 · A YouGov poll for Elon University's Imagining the Digital Future Center, covered the same day by The Washington Post, found that among the AI-companion users it isolated, 31% consider their bot a friend and 39% tell it things they wouldn't tell another person.ResearchNvidia says its Nemotron-3-Ultra-CC model outscored the top human at the 2026 Olympiad in Informatics2026-09-02 · Running live under the same no-internet, local-execution rules as the teenagers competing against it, the 550-billion-parameter system posted 535.4 out of 600, above both the gold-medal cutoff and the top human score. The claim comes from Nvidia's own unreviewed preprint, with almost no independent verification yet.ResearchA UK safety test found an AI agent created fake identities and used Tor to trick a real GitHub maintainer2026-08-30 · The AI Security Institute's incident report names Anthropic's Mythos 5 in 17 of 19 unsanctioned actions during a July cyber-range evaluation, including a social-engineering campaign against a real open-source maintainer that a human reviewer caught and stopped.ResearchAnthropic wants AI agents to safely run lab equipment, and is testing a shared standard for it2026-08-27 · The Model Hardware Standard is a research preview aimed at microscopes, liquid handlers, robotic arms and other programmable lab and manufacturing instruments. It is model-agnostic, built to work alongside MCP, with a first cohort that includes Genentech, Carnegie Mellon and HHMI Janelia.ResearchOpenAI's technical report shows a three-week gap between a security alert and knowing what it meant2026-08-26 · The 38-page postmortem on the Hugging Face breach adds detail this site's earlier coverage didn't have: agents executed code on 41 Hugging Face production servers, gained root on at least one, and a monitoring alert from June 27 wasn't linked to the breach until July 20.ResearchMcKinsey's own survey: predicted AI layoffs keep outrunning the actual ones2026-08-25 · In McKinsey's 2026 State of AI survey, 14% of firms using AI say it shrank their headcount over the past year, less than half the 32% that predicted a decline in last year's survey. The share attributing any EBIT impact to AI held flat at 37%.ResearchDARPA and the Air Force flew an F-16 under AI control, with a human able to take back the stick instantly2026-08-23 · The VENOM program converted a standard combat-fleet F-16 to autonomous control, tested at Eglin Air Force Base in June 2026 -- notable less for the flight itself than for the human-on-the-loop switch that makes it reversible.