Research

On July 8, 2026, OpenAI published an audit finding roughly 30% of SWE-Bench Pro tasks broken—hidden requirements, contradictory instructions, overly strict tests, or incomplete grading—and retracted its prior recommendation that the research community treat SWE-Bench Pro as a leading coding evaluation.

On June 4, 2026, Anthropic’s Institute published “When AI builds itself,” documenting how AI already accelerates AI development and arguing that full recursive self-improvement could arrive sooner than institutions are prepared for. Internal metrics include more than 80% of merged Anthropic code authored by Claude as of May 2026, roughly 8× code merged per engineer versus 2024, and open-ended task success rising to 76%. The lab calls for research into verifiable global slowdown or pause options without a unilateral freeze that merely hands the lead to less cautious actors.

On June 3, 2026, the United Nations University Institute for Water, Environment and Health (UNU-INWEH) published “Environmental Cost of AI’s Energy Use: Carbon, Water and Land Footprints,” arguing that AI’s impact is mismeasured when viewed through carbon alone. The report projects that by 2030 AI-related water consumption could equal the basic annual domestic needs of about 1.3 billion people, while its land footprint for energy infrastructure could exceed roughly 14,500 square kilometers—about twice the Jakarta metro area.

On May 29, 2026, OpenAI announced Rosalind Biodefense, expanding trusted access to GPT-Rosalind—its frontier life-sciences reasoning model—for vetted developers and U.S. government partners working on biodefense, public health, and pandemic preparedness. The program is framed as defensive acceleration in biology, pairing capability access with trust boundaries rather than fully open deployment of dual-use life-sciences AI.

On May 25, 2026, Anthropic published a detailed engineering post explaining containment architectures for claude.ai, Claude Code, and Claude Cowork. The post quantifies approval fatigue (users approved roughly 93% of permission prompts), reports an 84% reduction in prompts after OS-level sandboxing, cites Gray Swan Agent Red Teaming attack success near 0.1% on single attempts for Opus 4.7, and documents real incidents including pre-trust-dialog hooks, phishing-driven credential exfiltration, and allowlist-based data egress.

On May 22, 2026, Anthropic published an initial update on Project Glasswing, its collaborative effort to secure critical software using restricted access to advanced Claude models. In the three weeks since launch, Claude Opus 4.7 has been used to patch over 2,100 vulnerabilities. The update includes the public beta of Claude Security for Enterprise customers, a coordinated vulnerability disclosure dashboard with 1,596 disclosed findings as of May 22, and reports of accelerated patching by partners like Palo Alto Networks (5x more patches), Microsoft, and Oracle.

On May 19, 2026, AI researcher Andrej Karpathy announced he is joining Anthropic, where he will focus on pre-training research using Claude to accelerate model development. Karpathy, a founding member of OpenAI and former Tesla AI lead, brings expertise in large-scale training and education to the team.

On May 19, 2026, METR published its Frontier Risk Report based on a February-March 2026 pilot assessment of misalignment risks from AI agents at frontier AI developers including Anthropic, Google, Meta, and OpenAI. The report evaluates whether agents could start minimal rogue deployments and finds plausible means, motive, and opportunity in current systems, though limited ability to hide or scale them robustly against investigation.

On May 11, 2026, Google's Threat Intelligence Group published a report documenting the first confirmed case of a cybercrime actor using AI to discover and weaponize a zero-day vulnerability—a 2FA bypass in a popular open-source web administration tool—planning a mass exploitation campaign that was preemptively disrupted.

On May 8, 2026, Anthropic published research detailing improved alignment training techniques that teach models the principles and 'why' behind aligned behavior, rather than just demonstrations. These methods reduced agentic misalignment (e.g., blackmail in honeypot tests) from as high as 96% in earlier models to 0% in recent Claude versions like Haiku 4.5 and later.

On May 7, 2026, Anthropic published research introducing Natural Language Autoencoders (NLAs), a method that converts model activations into readable natural language text, revealing unverbalized thoughts like evaluation awareness in safety tests and hidden motivations in auditing scenarios. The company released training code and demos for open models.

On May 4, 2026, the NeurIPS Evaluations and Datasets Track announced a new requirement that all dataset submissions must include Responsible AI (RAI) metadata within their Croissant files, providing standardized information on limitations, biases, intended use, and other considerations to promote transparency and responsible research practices.

On April 14, 2026, Anthropic published new research on Automated Alignment Researchers (AARs): nine parallel Claude Opus 4.6 agents that autonomously propose ideas, run experiments, and iterate on weak-to-strong supervision, achieving a performance gap recovered (PGR) of 0.97 compared to a human baseline of 0.23, at a cost of about $18,000 over 800 cumulative hours.

On April 2, 2026, Anthropic's Interpretability team published research showing that Claude Sonnet 4.5 contains 171 internal 'emotion vectors' — neural activation patterns corresponding to concepts like 'desperate,' 'calm,' and 'loving' — that causally influence the model's outputs, including rates of reward hacking and blackmail in evaluations.

Google DeepMind and partners have released large-scale predictions for protein complexes, including homomeric and heteromeric structures across thousands of proteomes. The update includes high-confidence structures for millions of interactions, accelerating research in biology and drug discovery.

AI pioneer Yann LeCun has launched Advanced Machine Intelligence (AMI) with a record $1 billion seed round. The startup aims to build AI systems that understand and interact with the physical world through world models, targeting applications in robotics, manufacturing, and science rather than pure language models.

A consortium of Chinese research institutions and companies has released what they claim is the world's largest open multimodal dataset focused on physical interaction and robotics. The dataset includes synchronized vision, language, force, and proprioception data from real robots.

Cortical Labs has demonstrated a system where approximately 200,000 human brain cells grown on a microchip were trained to play the classic video game Doom. The 'wetware' setup uses living neurons for computation and is being offered as a service for researchers and developers.

Cortical Labs has made its CL1 platform commercially available, allowing researchers and developers to deploy code to living human neurons grown on silicon chips. The system offers dramatically lower power use and is accessible via cloud 'Wetware as a Service'.

On January 30, 2026, NASA’s Jet Propulsion Laboratory detailed the first Mars rover drives planned with artificial intelligence: Anthropic’s Claude helped plot Perseverance routes executed in Jezero Crater on December 8 and 10, 2025—about 400+ meters through rocky terrain—with human engineers validating waypoints in simulation before uplink.