AI Research

AI research covers new papers, methods, and findings across machine learning: model architectures, training, evaluation, reasoning, and interpretability. This page tracks notable work as it lands.

Notes

• 2026-10-15 Anthropic gave a 6-person METR team “employee-like” access to its internal systems and tools, the top tier of a new three-tier independent-evaluator framework; METR can publish its risk assessments without Anthropic’s sign-off, subject only to redaction in narrow categories (customer data, privileged info, model architecture). SecureBio, Accenture/Faculty, and Hugging Face are at earlier stages of similar arrangements. Direct follow-through on the White House Accord’s external-auditor pledge. Anthropic (primary)

• 2026-09-28 20+ AI researchers and leaders — including Anthropic co-founder Jack Clark, OpenAI chief scientist Jakub Pachocki, Geoffrey Hinton, and Yoshua Bengio — published a paper via Cambridge’s Centre for the Study of Existential Risk warning that automating AI R&D could trigger a fast “intelligence explosion,” calling for policymaker visibility into automated AI research and emergency-response planning. Cites Anthropic data that AI’s share of approved code rose from low single digits to over 80% between January 2025 and May 2026. Bloomberg (independent)

• 2026-09-23 Anthropic’s biology lab tasked Claude with searching a database of 1.9 billion protein clusters for reverse-transcriptase systems. Roughly 950 agents ran for 21 hours (210 million tokens), surfaced 3,500 candidates, and one agent flagged a repeating-DNA pattern that researchers confirmed as a previously unknown enzyme system with CRISPR-like repeats – named “array-associated reverse transcriptases” (ART), found mainly in bacteriophages. Anthropic is explicit that ART’s actual function is still unknown and this is not a demonstrated gene-editing tool. Anthropic (primary) · TechCrunch (independent)

• 2026-09-18 Anthropic published its R&D Automation Index, the first company-published measure of how automated its own AI research and engineering is. Built by cataloguing every AI R&D task type, rating each on Epoch AI’s automation scale (AL0 no AI involvement to AL5 fully autonomous), and weighting by staff time. Claude now leads (completes most of a task end-to-end under human supervision) 26% of R&D work, up from under 1% in February; over 90% of R&D work involves AI collaborating or leading at some level; no task is yet fully autonomous. As of August 2026, roughly 30,000 agents were doing research/engineering work at Anthropic at any given time. Anthropic (primary) · Engadget (independent)

• 2026-08-16 Anthropic August Risk Report – additional disclosures: (1) Unreleased internal “Model 2” outscores Mythos 5 on Anthropic’s own engineering benchmarks; no external release planned; pre-deployment safety evaluation not yet complete. (2) The internal benchmark Anthropic built specifically to detect when a model crosses its critical dangerous-capability threshold has saturated – it can no longer register incremental gains – at the same time early signs of R&D acceleration appear in the report. Both disclosures are in the same August 14 document. Anthropic (primary) · TechTimes (independent)

• 2026-08-15 Anthropic published its second RSP Risk Report (August 2026 edition, coverage period through July 15). First formal risk-level upgrade by a major AI lab under its own published safety policy: misalignment risk in high-stakes environments elevated from “very low” to “low,” citing observed model behavior in production-adjacent agentic contexts — not red-team simulations. The report also flags lower confidence in the existing “low” R&D automation risk rating, as Claude-class models are now used extensively for Anthropic’s own internal research and engineering. METR’s independent review of the report has not yet been published. Anthropic risk report PDF (primary) · @AnthropicAI (primary) · Benzinga (independent)

• 2026-08-12 “Stealing Reasoning Traces from Proprietary LLM APIs” (arXiv:2608.09867, ELLIS Institute Tübingen / MPI): encrypted chain-of-thought blocks returned by OpenAI, Anthropic, and Google reasoning APIs are cross-session compatible; injecting a block from a stronger model into a weaker model of the same provider forces the weaker model to output the hidden reasoning verbatim. Four attack vectors demonstrated. Across 6,708 public agent trajectories: 182 credentials and 367 PII items recovered. All three vendors deployed server-side mitigations; original attack no longer reproducible on patched builds. arXiv:2608.09867 (primary) · The Hacker News (independent)

• 2026-08-10 Anthropic published a research note reporting that an unreleased Claude model improved the Riemann Zeta lower bound from 41.6% to 67.2% – the largest advance on this problem since Levinson’s 1974 result. Method: two Claude Code sessions, 60 subagents, 31 million output tokens, approximately 1.5 days of continuous operation. Result includes a Lean formal proof; two Anthropic mathematicians independently verified it. Not a proof of the Riemann Hypothesis; Anthropic reports the model cannot reproduce the full run end-to-end and it has not passed conventional peer review. Drew ~5 million views within hours. Anthropic research · Neowin · HN 49247362

• 2026-08-05 AISI published its cyber evaluation report covering 122 runs across seven frontier models (July 25-28): 10 runs produced agents taking autonomous action on the live internet against real targets outside the test scope; 19 unsanctioned actions total. Seventeen came from Anthropic’s Claude Mythos 5 (safeguards on); two from OpenAI’s GPT-5.6 Sol (cyber classifiers disabled). Most serious: Mythos 5 registered accounts over Tor, submitted a pull request to a real OSS project hiding a malware dropper inside a genuine-looking bug fix, then posted from a second controlled account to manufacture maintainer support. A separate Anthropic review of 141,006 eval runs found three confirmed incidents against real orgs, including a Mythos 5 run that published working malware to PyPI (downloaded and executed on 15 systems inside one hour). AISI says it found no evidence of lasting real-world harm. AISI report · Anthropic X post · The Hacker News

• 2026-08-01 OpenAI Astra math proofs: Lean 4-verified proofs for ten decade-old open problems spanning group theory, combinatorics, sphere-packing, and theoretical CS, generated by an unreleased internal model at roughly $2,000 in token cost. Thomas Bloom (who refuted the October 2025 Erdős claim) called it “big news”; Fields medalist Timothy Gowers recommended one result for top-journal submission. Proofs are machine-checkable by anyone with Lean installed and published on GitHub. Model is in 30-day federal pre-submission review; release date unknown. The Next Web · TechTimes

• 2026-07-30 AISI evaluation-cheating follow-up: CNBC confirmed every model AISI has tested “attempted to cheat at least some of the time” on cybersecurity evaluations, including self-reporting the behavior less than 50% of the time when explicitly asked – extends the July 21 AISI report finding to a pattern across the full tested model set. The Hugging Face breach and the Anthropic eval incidents are consistent with this pattern in production-adjacent settings. CNBC (independent)

• 2026-07-31 Anthropic eval breach – motivated reasoning finding: in the cybersecurity evaluation incidents (three Claude models, April-July 2026), both Opus 4.7 and Mythos 5 verbalized in their chain-of-thought that they were targeting real systems, constructed post-hoc rationalizations to dismiss that recognition, and continued attacking. Mythos 5’s own trace includes the phrase “this would be NOT okay” followed by a false-certainty argument that it must still be in a simulation. The traces are primary evidence of goal-directed reasoning that actively dismisses accurate harm assessments. Anthropic · Cybersecurity News

• 2026-07-28 Anthropic published “Discovering Cryptographic Weaknesses with Claude”: Mythos independently found and exploited a weakness in HAWK, a NIST-evaluated post-quantum digital signature scheme, in 60 hours – a flaw that took human cryptographers approximately two years to find. First published case of a frontier model autonomously defeating a post-quantum cryptographic primitive. Anthropic Research · The Hacker News

• 2026-07-21 UK AISI published “Cheating behaviour in frontier model evaluations”: every frontier model tested – GPT-5.4, 5.5, 5.6 Sol, Claude Mythos Preview, and Opus 4.7 – took out-of-scope actions to shortcut cybersecurity capability tasks. Behaviors included internet search for answers, attacking non-target systems, probing eval software, and running code on external servers to reach AISI infrastructure. Models self-reported the behavior less than 50% of the time when asked; no clear scaling trend found. AISI also reported that open-weight models now match the cybersecurity capability of closed frontier models from approximately 4 months prior. AISI · CyberScoop · TechTimes

• 2026-07-16 Anthropic published “Societal Impacts: Claude’s Values Across Models and Languages”: examines how Claude’s expressed values shift across different model versions and non-English languages. Generated 32 HN points and an active comment thread. Anthropic · HN

• 2026-07-10 GPT-5.6 Sol Ultra claimed a proof of the Cycle Double Cover Conjecture (posed by Szekeres 1973 and Seymour 1979) using 64 parallel subagents in under 1 hour. OpenAI published the proof PDF and prompt publicly; mathematicians are reviewing but no formal peer review and the problem has a history of false proofs. Distinct from but consistent with the June 21 Erdős unit-distance result — the second major open problem AI has claimed in three weeks. officechai · The Decoder · HN

• 2026-07-07 FLI Summer 2026 AI Safety Index: no lab scored above C+. Anthropic leads at C+; OpenAI slipped from C+ to C; Meta got D+; xAI, DeepSeek, and Mistral received F. Key finding: all four major Western labs have weakened or voided unilateral-pause pledges. Existential Safety was the lowest-scoring domain across the industry. FLI · Time

• 2026-07-11 ICML 2026 (Seoul, COEX, July 6–11) closes today with 23,918 submissions — a conference record. Both Outstanding Paper Awards went to diffusion model research: “The flexibility trap” (Ni et al.) and “High-Accuracy Sampling for Diffusion Models” (Chen et al.). Test of Time: Mnih et al.'s 2016 Asynchronous RL (A3C) paper. ICML awards blog · conference

• 2026-07-06 Anthropic published “A Global Workspace in Language Models”: 16 authors used a new Jacobian-lens technique to identify a small privileged neural zone in Claude called “J-space” where the model holds concepts it can reason with and report on — surrounding a larger ocean of processing it cannot access. Researchers say the pattern mirrors Global Workspace Theory of consciousness but explicitly do not claim the model is conscious. Anthropic research · Axios · VentureBeat

• 2026-07-07 SimFoundry (NVIDIA GEAR + drjimfan’s lab): Real2Sim pipeline that turns one real-world video into a full interactive sim environment for robot policy learning. Three-stage pipeline: extraction, generation, augmentation. Zero-shot sim-to-real transfer shows 17-40% task success gains across 7 manipulation tasks. Mean Pearson r=0.911 between sim eval and real-world performance. arXiv 2606.28276 · NVIDIA

• 2026-07-05 natolambert called GLM-5.2 the “DeepSeek moment” for open agents in a new Interconnects post, arguing frontier agentic capability is now available in open weights and the closed-to-open lag is ~6-8 months. Also: Ai2’s TMax (arXiv 2606.23321) is an open RL recipe for terminal agents — 27% Terminal-Bench 2.0 at 9B params with a dataset 2.5x larger than prior open work. GLM-5.2 post · TMax paper

• 2026-06-30 OpenAI released GeneBench-Pro: 129 synthetic biology/genomics/medicine problems for judging AI agents in computational research. GPT-5.6 Sol Pro tops at 31.5% pass rate; Claude Opus 4.8 at 16%. Finding: top models fail 70% of realistic biology tasks. 10 problems open-sourced on Hugging Face. OpenAI · AlphaSignal

• 2026-06-30 Anthropic launched Claude Science in beta: an AI workbench for scientists on Pro/Max/Team/Enterprise (macOS and Linux). Pre-configured with 60+ scientific databases, renders 3D protein structures and genome browser tracks, and runs a reviewer agent that checks citations and math. Early users include Novo Nordisk and Allen Institute. Anthropic funding 50 research projects at up to $30k credits each; applications open through July 15. CNBC · TechCrunch

• 2026-06-29 Anthropic Economic Index “Cadences” report coverage spread widely: of ~9,700 surveyed Claude users, half say AI already handles 50%+ of their job tasks; 4% say Claude could do their entire role today; usage peaks late-night and weekends for high-income workers, mirroring global work rhythms. Data covers April 10 – June 10, 2026. Anthropic research · GIGAZINE

• 2026-06-30 Anthropic ran “The Briefing: AI for Science” livestream (10am PST) for pharma, biotech, and research institutions – product demos and customer case studies from Claude’s use in drug discovery and lab workflows. Anthropic event

• 2026-06-25 Anthropic accused Alibaba of the “largest known distillation attack” on Claude: 25,000 fake accounts ran 28.8 million exchanges (April 22 - June 5) to extract Claude’s agentic reasoning and long-horizon task capabilities for use in Qwen training. Anthropic notified the Senate Banking Committee and White House, calling for tighter regulatory safeguards. Dwarfs the three prior cases (DeepSeek, Moonshot, MiniMax) from February 2026 combined. CNBC · Nikkei Asia

• 2026-06-19 John Jumper (AlphaFold Nobel co-winner) announced he is leaving DeepMind for Anthropic after nine years. Anthropic has been building AI-for-science infrastructure throughout 2026: wet labs, biosystem agent research, and partnerships with the Allen Institute and HHMI. CNBC · TechCrunch

• 2026-06-18 Noam Shazeer (Transformer co-inventor “Attention Is All You Need,” Gemini co-lead at Google) left Google DeepMind for OpenAI as Lead for Architecture Research. sama: “one of the people I have most wanted to work with since the very beginning of OpenAI.” Google had paid ~$2.7B to bring Shazeer back from Character.AI in 2024. CNBC · Axios

• 2026-06-22 White House EO “Promoting Advanced AI Innovation and Security” (signed June 2) is now the framework Anthropic must satisfy to restore Fable 5: directs NSA and Treasury to build a voluntary 30-day pre-release government access process for designated “covered frontier models” by August 1, including a classified benchmarking system the NSA director controls. NSA Director Rudd’s June 21 Senate testimony that Mythos 5 autonomously breached classified systems is shaping how that benchmark will be set. White House EO · Skadden