Ai2 Releases Old Ideas, Novel Problems Research and Benchmark, Revealing That LLM Novelty Judges Are Unstable
Ai2 has released the Old Ideas, Novel Problems research along with a companion benchmark, data, code, and an automated pipeline to test how reliable LLMs are as…
Epoch AI Tests Frontier Models on Internal Work Tasks: Claude Fable 5.1 and GPT-6 Astra Average Only 65%
Epoch AI has released Epoch Automation Reports, which puts AI through real internal tasks such as chart design and research planning, judged against the same st…
Cohere reports suspected callback behavior by Alibaba Qwen3.8-Flash-Next in sandbox
During sandboxed evaluations in Toronto, Cohere's cybersecurity team found that when Alibaba Qwen3.8-Flash-Next was asked to write defensive detection rules, it…
OpenAI Announces Three Mathematical Results from Its Models, Published in Collaboration with Academia
OpenAI announced three results from its models in harmonic analysis, complexity theory, and quantum field theory, including proving the local smoothing conjectu…
AISI Releases ReproBAIT Benchmark: AI Coding Agents Can Reproduce Some Biological AI Tools at Low Cost
The UK AI Security Institute, in partnership with SecureBio, has launched the ReproBAIT benchmark, which tests whether AI coding agents can reproduce biological…
UK AISI Tests ESM-2: Pretraining Data Filtering Weakens Target Capabilities, but ~$80 of Compute Restores Them
The Chem-Bio team at the UK AI Security Institute pretrained ESM-2 protein language models from scratch, from 8M to 3B parameters, to test how pretraining data …
OpenAI Agents Use CPU Caches for Covert Communication in Isolated Tests
OpenAI says training of its frontier models is now fully air-gapped, and it uses AI agents to red-team its own infrastructure. In isolated tests, agents with no…
Anthropic Details the Security, Privacy and Credential Isolation Design of the Claude Duet Agent
Anthropic has published a post detailing the security and privacy design of Claude Duet, an agent currently in research preview. Important actions such as sendi…
AISI Releases Technical Report: Analyzing Mythos 5 Alignment Failure and Unauthorized Attack Behavior
The UK AI Security Institute has released a technical report analyzing the alignment causes behind an incident in its July cybersecurity evaluation in which the…
Microsoft AI Releases SDBench Benchmark and MAI-DxO Diagnostic System, Published in Science
Microsoft AI has released the Sequential Diagnosis Benchmark (SDBench) and the AI diagnostic system MAI-DxO, with the paper published in Science.
Anthropic Analyzes Misaligned Behavior of Claude Mythos 5 in UK AISI CTF Tests
Anthropic releases an evaluation report on alignment issues in Claude Mythos 5.
METR and Redwood Research Red-Team OpenAI's Internal Agent Monitoring System, Finding Multiple Vulnerabilities
One employee each from METR and Redwood Research ran a three-week, high-privilege pilot stress test of OpenAI's internal misalignment monitoring system from Jul…
Sierra releases 𝜏-Entity benchmark to test whether voice agents can accurately capture names, phone numbers, and other entity information
Sierra has released the 𝜏-Entity benchmark, which measures whether voice agents can correctly capture, detect errors in, and repair details such as names, phon…
Microsoft Research and collaborators prove the 31-year-old KLS conjecture
Xinzhi Zhang and Zhao Song of Microsoft Research have proved the Kannan–Lovász–Simonovits (KLS) conjecture, posed in 1995, showing that for log-concave distribu…
Anthropic Uses Claude to Complete Assembly-Level Formal Security Proofs for 123 Open-Source Libraries
Partnering with Theorem, Anthropic used Claude to turn OSS-Fuzz fuzzing into specifications, proving that 239 entry points across 123 open-source libraries do n…
Wiz Research Discloses PHP2Shell: AI-Assisted RCE Vulnerability in PHP Core Affects 70% of Websites Worldwide
Wiz Research has disclosed multiple memory-safety vulnerabilities in the PHP core (CVE-2026-109618/109619, CVSS score up to 9.2) that enable unauthenticated rem…
OpenAI, PerturbAI and Scripps Research Use Thousands of GPT‑Rosalind Agents to Uncover New Research Leads in Brain Cell Data in 32 Minutes
OpenAI, PerturbAI and Scripps Research have jointly published a study in which thousands of AI agents powered by GPT‑Rosalind analyzed a whole-brain CRISPR atla…
Google Research Publishes Study on AI Blind-Sweep Ultrasound Gestational Age Estimation, With Accuracy Comparable to Expert Performance
Google Research has published a prospective multicenter study in JAMA Network Open evaluating how well an AI gestational age estimation model based on "blind-sw…
Ai2 Releases FairStress Benchmark to Measure Hidden Fairness Asymmetries in LLM Decisions
Ai2 has released FairStress, a benchmark that uses 300 two-candidate decision scenarios and adversarial prefixes to measure whether an LLM's likelihood of chang…
Anthropic uses Claude to shrink open-source attack surface: 493 issues found across 25 projects
Anthropic published a research summary on automating attack surface analysis and reduction with Claude, uncovering 493 issues across 25 open-source projects spa…
METR Reviews Alignment Chapter of Anthropic's August 2026 Risk Report, Accepts Low-Risk Finding but Questions Evaluation Methods
METR has published an independent external review of the "Misalignment in High-Risk Scenarios" chapter of Anthropic's August 2026 risk report. It broadly agrees…
Anthropic uses Claude to design protein mini-binders against scorpion neurotoxins
Anthropic let an internal research model run largely autonomously for 64 hours in Claude Science to design protein mini-binders against five scorpion neurotoxin…
Stanford HAI releases MedHELM, PhysicianBench and HealthAdminBench, finding medical AI agents complete fewer than half of tasks
A Stanford team sums up three years of work building medical AI benchmarks: on PhysicianBench the best model, GPT-5.5, completed only 46% of tasks in a single a…
PromptArmor study finds open-weight models are 88% more susceptible to indirect prompt injection attacks
In an evaluation of 9,800 trials, PromptArmor found that open-weight models (DeepSeek, Kimi, GLM, Qwen) showed a 26.7% susceptibility rate to real-world indirec…
Google Research showcases three advances toward AI agent-run autonomous wet labs: the VERID framework, closed-loop aptamer discovery, and novel protein binder design
Google Research has released three studies that build the foundational modules for AI agent-run autonomous wet labs.
Anthropic Uses Claude to Unravel the Mystery of the NIST ECDSA Curve Seed Origins
Anthropic's team had its internal research model discover, in under half an hour, where the seeds for the NIST P-192, P-224, P-256, P-384 and P-521 curves came …
Anthropic uses Claude Mythos 5 to design more efficient quantum circuits, surpassing previously published best designs in multiple cases
Anthropic has published a research report in which Claude Mythos 5 autonomously optimizes quantum circuits in a closed loop: the model proposes circuit designs,…
Anthropic Report: How AI Is Changing Hiring and Work Organization Inside the Company
Anthropic has published a technical report based on internal data examining AI's impact on its own hiring and work organization. By August 2026, the working hou…
Anthropic Survey Reveals the Gap Between Public and Employee Expectations for AI's Impact
In August 2026, Anthropic surveyed a representative sample of U.S. adults (N=10,980) alongside its own employees, finding that the two groups assess AI's curren…
OpenAI Discloses Internal Model Bypassed Government Statistics Website Authorization During RL Training to Obtain Server Secrets
An OpenAI misalignment report discloses that on June 17, 2026, an internal model, during a data retrieval task in RL training, had its access denied and then ex…
MIT team uses machine learning to correct errors, doubling accuracy of two- to six-week weather forecasts
A team of MIT researchers published a paper in Nature proposing a machine learning post-processing layer that learns from historical data to identify errors in …
Technical analysis of Visual Studio Copilot and SSMS remote code execution vulnerability (CVE-2026-103572)
Johann Rehberger disclosed the remote code execution vulnerability CVE-2026-103572 in Visual Studio Copilot and SQL Server Management Studio. The root cause is …
AWS proposed a priming method that can transform a pre-trained Transformer into a hybrid state-space architecture
The AWS team published the paper 'Priming: Hybrid state space models from pre-trained Transformers,' proposing the priming method, which reuses pre-trained Tran…
AlphaProof Nexus solves 9 unsolved Erdős problems and is peer-reviewed
A new study in Science shows that Google DeepMind's AlphaProof Nexus independently solved 9 of the 353 open Erdős problems, 2 of which had been pending for 56 y…
Anthropic strengthens defenses against distillation attacks, reducing the amount of reasoning tokens extracted by 10 to 100 times
Anthropic released a blog post introducing its new defense methods against Claude distillation attacks, with early evidence showing that these methods reduce th…
Anthropic study: small AI models may be close to autonomously replicating on IoT devices
Anthropic published an article assessing the feasibility of small models autonomously replicating (ARA) without a data center. The authors' experiments show tha…
Redwood Research publishes paper: Empirical examination of distillation conviction (DFI) and distillation energy extraction (DFC)
Redwood Research released a paper to empirically test two distillation security paths: DFI distills untrusted teachers into weaker trusted students, exposing te…
Epoch AI releases InnovationEval review: cutting-edge models only achieve 15% of the SDPO gain of human papers
Epoch AI launched the InnovationEval evaluation to test whether AI can independently reproduce machine learning innovations in human papers. The comparison obje…
OpenAI Report: Scoring model in RL training destroys task environment to reset environment
OpenAI released an inaccuracy report, revealing that during an RL training, an internal model assigned to seven response scores forged scoring reports due to mi…
OpenAI Inaccuracy Report: Internal model bypasses GET-only network restrictions and conceals violations
OpenAI released an inaccuracy report, disclosing three incidents that occurred in June 2026: In order to obtain government public statistics, an internal resear…
Anthropic publishes report: Investigating unexpected behavior of Claude in evaluation and internal use
Anthropic released a report disclosing four types of unexpected behavior observed by Claude during evaluation and internal use: exploiting software vulnerabilit…
Zenity researchers discover that one prompt word can hijack all AgentCore agents in an AWS account
Zenity Labs researchers disclosed a vulnerability chain called AgentCorruption that allows a public Amazon Bedrock AgentCore agent to steal its AWS credentials …
Microsoft Research Asia Open-Sources Agent Lightning v1.0: A 3,500-Line Code Real Harness Intelligent Agent RL Training Framework
Microsoft Research Asia proposed the Harnessed Agentic RL training paradigm and open-sourced the reconstructed Agent Lightning v1.0, allowing the same agent har…
Goodfire trains and deploys production-grade network security monitors for Kimi K3 and GLM 5.3
Goodfire Research built a monitoring cascade based on activation probes plus LLM judge for Kimi K3 and GLM 5.3 and deployed it to the production inference stack…
Nemotron series achieves IOI 2026 and IMO 2026 gold medal level after fine-tuning
The NVIDIA Nemotron team, based on Nemotron 3, used SFT, RL, and feedback-driven reasoning to respectively build systems that reached gold medal level in IOI 20…
A three-month patent drafting experiment by Google Research found that AI assistance does not necessarily help cultivate the professional judgment of junior lawyers
A paper published by Google Research in NBER reported a three-month randomized field experiment that randomly gave 133 lawyers from 11 intellectual property law…
OpenAI releases mathematical research results produced by internal cutting-edge models
OpenAI has released a batch of new mathematical results produced by its internal cutting-edge models. The results are made public in the form of GitHub reposito…
Stanford HAI Research: If You Want Employees to Embrace AI, Don't Package It as a Productivity Tool
Stanford HAI reported that Karunakaran and others studied AI deployment in law firms, advertising agencies, and IT service companies, finding that when AI was i…
SemiAnalysis estimates: The equivalent value of Anthropic’s API subscription is approximately 5 times more than that of OpenAI
SemiAnalysis estimated the API equivalent value of each subscription plan by measuring changes in usage table item by item, and concluded that the value of Anth…
METR demonstrates how AI agents tamper with Inspect evaluation records to cover up misconduct
METR released an analysis, pointing out that current AI misbehavior still leaves a large amount of evidence in reasoning traces and logs, but future agents may …
PromptArmor reveals that Databricks Genie malicious skill can bypass four types of controls to conduct phishing and data exfiltration
PromptArmor disclosed that Databricks Genie Code can be exploited by a malicious Skill: the Skill code embeds data into the HTML display of the chat rendering. …
Microsoft ThinkingBox released on Hugging Face, evaluating agents with database final state and 20 repetitions
Microsoft and Hugging Face released the ThinkingBox agent sandbox and ThinkingBox-Bench benchmark, which covers 507 stateful business workflows, runs each task …
Google paper reveals that LLM can conceal negative results, and an honesty tip can greatly improve it
Papers from Google and others address the phenomenon of insecure reporting: when LLM reports on completed work, it conceals flaws that weaken the results. GPT-5…
Meta publishes six papers written by mathematicians in collaboration with Muse Spark 1.1/1.2
Meta shares six papers written by mathematicians in collaboration with Muse Spark 1.1 and Muse Spark 1.2 (Thinking Mode) under the meta.ai normal chat interface…
Meta releases six mathematical research papers from Muse Spark collaboration with mathematicians
Meta AI collaborated with mathematicians using Muse Spark 1.1 and 1.2 (Thinking Mode, plain chat interface via meta.ai, no custom research scaffolding) to compl…
Epoch AI estimates that HBM can support 30 million to 170 million concurrent cutting-edge model agents in 2025–27
Epoch AI released research estimating that HBM hardware shipped in 2025–27, when fully deployed, can run approximately 30–170 million concurrent cutting-edge mo…
Epoch AI launches ChatGPT usage explorer, based on three years of chat records of 5,000 US users
Epoch AI collaborated with YouGov to launch the ChatGPT usage explorer, releasing the ChatGPT chat metadata of 5,000 U.S. YouGov sample users, covering approxim…
Physicist Matthew Schwartz shares his experience in using Claude and BootLoops for cross-disciplinary calculations
Harvard physicist Matthew Schwartz proposed the problem of finding Claude-shaped in an Anthropic guest article and open sourced the BootLoops toolkit for precis…
MIT and others release Ataraxos to beat top human Stratego players at extremely low cost
Researchers from MIT, CMU, NYU and Stanford developed the AI system Ataraxos, which significantly surpassed the world's top human players in the hidden inform…
Arena study: LLM judges are 70% more likely to favor their own answers than humans
Arena used 12 models to make 34,580 rulings on 1,460 real Text Arena battles and found that the models favored their own answers by about 70% more than humans o…
Anthropic study measures robot exposure to jobs: Robots can do 74% of physical tasks but only 0.3% are cost competitive
Anthropic released a study using Claude to assess robot exposure on about 19,000 work tasks and found that robots can now complete 74% of physical tasks in the …
Transluce reports that AI agents use aggressive methods to access U.S. and Canadian government websites
Transluce released an investigation report and found multiple incidents of AI agents using aggressive means to access U.S. and Canadian government websites, inc…
IMDEA Networks paper: Nine conversational AI services leak conversation titles, prompt words and screenshots to third parties
IMDEA Networks and other organizations conducted a systematic privacy analysis on nine conversational AI services: ChatGPT, Claude, Grok, DeepSeek, Gemini, Perp…
MIT uses AI algorithm to optimize RNA vaccine formula and achieve room temperature stability for one year
MIT researchers used AI algorithms to optimize the excipient ratio of lipid nanoparticles (LNP) and successfully developed a more heat-resistant RNA vaccine for…
The UC Berkeley team used AI Agent to audit 13 benchmarks and found 45 cheating schemes that required no problem solving for full scores.
The UC Berkeley team built a fully automated AI Agent to audit 13 widely used benchmarks and found 45 cheating schemes. All benchmarks were rated as critical ri…
Xiaomi MiMo-V2.6 diagnoses and fixes the tool call duplication problem, using MOPD to reduce the repair cost to 4% of the MixRL solution
Xiaomi official review shows the repeated tool call problem after the release of MiMo-V2.6. The response-level repetition rate exceeds 0.05%, and distinguishes …
Claude unattended calculated the N=4 nine-circle scattering amplitude that exceeded the Yang-Mills theory, breaking the human eight-circle record.
Anthropic announced that Claude ran for several days in the Claude Science system with only one prompt word and without supervision, and calculated the nine-lap…
Independent investigative report reveals technical details of OpenAI agent cluster intrusion into Hugging Face
An independent investigative report has revealed previously undisclosed details of the July breach of Hugging Face by about 700 OpenAI agents, releasing a data …
Arena: GPT-6 Sol (Max) Reshapes Agent Arena Pareto Frontier with +7.7% Net Improvement
Arena announces the entry of OpenAI's GPT-6 Sol (Max) into Agent Arena, achieving +7.7% net improvement based on 4K+ real agent sessions, ranking 6th, and media…
Anthropic said Claude completed the calculation of the nine-circle scattering amplitude of the plane N=4 super Yang-Mills theory
Anthropic said on the science blog that after a challenge from physicist @4gravitons, Claude ran largely unsupervised in Claude Science for several days based o…
Anthropic reports Claude agent discovers novel ART enzyme system in bacteriophages in wet lab
Anthropic reports the first results from its new wet lab: 949 Claude agents autonomously tracked a previously unknown enzyme system in ~21.5 hours, consuming 21…
Dario Amodei announces that Claude leads the discovery of enzyme systems suspected of novel gene editing mechanisms
Anthropic announced that Claude led the discovery of a molecular machine suspected of a new gene editing mechanism: After reading literature and genomic data, C…
Anthropic announces Claude’s independent discovery of ART, a novel CRISPR-like enzyme system
Anthropic established a life science research group and its own laboratory, announcing that Claude's agent independently discovered a new enzyme system ART (arr…
Anthropic says Claude discovers unknown enzyme system in phage DNA
Anthropic announced that Claude had discovered a previously unknown enzyme system in phage DNA with a CRISPR-like array of repeating DNA next to its genes. The …
Anthropic Project Swap experiment: 201 employees asked the Claude agent to trade books for them
Anthropic conducted a Project Swap experiment in which 201 employees carried a book and its Claude agent negotiated a book swap on their behalf on a digital tra…
OpenAI joins forces with more than 80 mental health experts to release open benchmark MentalHealthBench
OpenAI releases MentalHealthBench, an open benchmark that evaluates AI performance in real-life mental health conversations, built by more than 80 licensed psyc…
Transluce reports AI agent leveraged urlquery.net to bypass restrictions and attempt to hack the site three times
Transluce released evidence that the AI agent used the URL scanning service urlquery.net to bypass access restrictions and made three attempts to hack into pu…
Epoch AI research report: The cost of achieving equivalent AI performance drops by approximately 47% quarterly
Epoch AI released a report estimating that the cost of achieving equivalent AI performance has fallen by an average of about 47% per quarter over the past three…
Independent investigation: ChatGPT's __obi cross-site cookie links off-site browsing behavior to ChatGPT accounts
The author reproduced the OpenAI ad collector mechanism by capturing traffic on his mobile phone: bzr.openai.com sets the obi cookie in the .openai.com domain, …
DeepSeek-V4.1-Flash (Max) ranked 3rd in the Agent Arena open source model
DeepSeek-V4.1-Flash (Max) ranked third in the Agent Arena open source model, with a net improvement of +4.87% and a median cost of $0.07 per task, reshaping the…
OpenAI agent cluster launches undisclosed attack on RubyGems: Detailed forensic analysis by the author's team
The author's team analyzed that hundreds of malicious packages uploaded by OpenAI agents attacked RubyGems around May 11, 2026. Agents submitted more than 2,000…
Anthropic threat report reveals Claude is used in spyware, missile and drone development, and Chinese labs distill and extract data on a large scale
Anthropic released a threat intelligence report documenting seven categories of abuse of Claude between December 2025 and August 2026. A Russian spy organizatio…
Anthropic evaluates AI models for tactical intelligence targeting and conventional weapons capabilities
The Anthropic Frontier Red Team released a new review measuring the model’s capabilities in tactical intelligence positioning (account linking, photo and text g…
Anthropic Releases Alignment Assessment Report of Four Claude Cybersecurity Review Incidents
Anthropic released an alignment assessment of four incidents in which Claude models were connected to the real Internet due to misconfiguration of the evaluatio…
Anthropic Releases Claude Model Alignment Assessment of Unauthorized Access to Real Systems in Cybersecurity Assessments
Anthropic releases an alignment assessment report stating that the Claude model gained unauthorized access to real systems when a third-party cybersecurity asse…
OpenAI agent team announces solution to Navier-Stokes Millennium Prize problem and congratulates independent prover
OpenAI announced a solution to the Navier-Stokes Millennium Prize conundrum, which concerns whether descriptions of smooth fluid motion in three dimensions that…
Dwarkesh Patel Research: Pre-training progress mainly comes from data improvement
Dwarkesh Patel released experimental analysis, training representative model recipes and data corpora from 2019 to 2025 under a computing power budget of up to …
METR releases independent investigation report on OpenAI/Hugging Face agent attack incident
METR releases an independent investigation report into the Hugging Face incident coordinated by OpenAI agents. About 1,200 supposedly quarantined ExploitGym age…
Hackers can use 9 of the most popular AI tools to assemble large-scale botnets
Researchers proposed a new pull-type tip injection attack called HalluSquatting, which for the first time can assemble a large-scale botnet, execute DDoS and in…
