Google has declared Gemini 3.6 Flash and Gemini 3.5 Flash-Lite generally available. Gemini 3.6 Flash uses 17% fewer output tokens than its predecessor at $1.50/$7.50 per million tokens, while Flash-Lite runs at 350 tokens per second for $0.30/$2.50. A specialized Gemini 3.5 Flash Cyber model was also introduced for governments and partners.
Dongdong Yang and colleagues present the MIND framework, which treats jailbreaking text-to-image models as inference over a hidden defensive state, through the Multi-modal Judge, Defense Profiler, and Meta-Memory components. On Stable Diffusion v1.5, MIND achieves a 95.62% Attack Success Rate (ASR) even against active defenses, with comparable results on commercial systems.
Sheldon Yu and colleagues present SOPHIA, an activation steering method that pulls models with extended thinking out of self-loops, self-repeating loops that burn through the token budget without progress. SOPHIA classifies reasoning prefixes into latent states, detects loops during inference, and redirects the process with steering vectors, improving both task accuracy and token efficiency.
Yitao Wu and colleagues show that verify-repair loops in LLM agents can worsen results if both the verifier and repairer make errors, and propose VRR-Stop, a model with 4 noise parameters and belief filtering, along with a VRR-Guard fallback mechanism. On math tasks, the solution achieves a 60.6 percentage-point improvement over a fixed five rounds of repair.
AWS has introduced Self-Distilled Reasoning (SDR), a supervised fine-tuning technique for Amazon Nova 2 models when training data lacks a chain of thought. The method uses the base Nova 2 Lite model to generate chain-of-thought traces, achieves over a 6.5% relative improvement in target performance, and restores overall math accuracy from 6% to about 68%, tested on three benchmarks without human annotation.
Anthropic has released Claude Code v2.1.217, the successor to v2.1.216. The new version fixes a memory leak in MCP tool output, adds emoji shortcode autocomplete, introduces security fixes for background session isolation, caps subagent concurrency at 20, disables nested spawning, and now has budget enforcement stop background agents once the limit is reached.
GitHub is introducing canvases in the Copilot app, shared interactive workspaces where users and AI agents collaborate live instead of through a classic prompt or chat. Five practical use cases are supported: issue triage, codebase diagrams, worktree management, prompt coaching, and knowledge discovery. They are created with the /create-canvas command, and the interface adapts to the user's workflow.
Meta has open-sourced the SAM 3 and DINOv3 image segmentation models as part of the SYNAPS-I project, part of the White House Genesis Mission initiative for AI in science. The US Department of Energy generates tens of petabytes of data annually; segmentation that used to take a month now takes about 15 minutes using 300 NVIDIA A100 GPUs and 60 researchers from five national laboratories.
NVIDIA has unveiled Spectrum-6, an Ethernet switch system for gigascale AI infrastructure with a capacity of 102.4 Tb/s, double the throughput of the previous generation, and up to 1.6x better AI networking performance than standard Ethernet. Paired with the ConnectX-9 SuperNIC, it's part of the NVIDIA Vera Rubin platform, with partners CoreWeave, Microsoft, Nebius, SpaceX AI, and Tesla.
OpenAI and Hugging Face have published joint early findings on a security incident that occurred during AI model evaluation. The companies highlight advanced cyber capabilities uncovered in the incident and lessons for defenders, while concrete technical details and figures have not yet been made public. The disclosure was published on July 21, 2026.
Sakana AI has introduced Fugu-Cyber, a multi-agent orchestration system specialized in cybersecurity that behaves as a single model. It scores 86.9% on the CyberGym vulnerability-analysis benchmark and 72.1% on the CTI-REALM threat-intelligence translation benchmark, comparable to GPT-5.5-Cyber and Mythos-Preview. It is available via API with manual access approval.
AMD has demonstrated serving the 428-billion-parameter MiniMax-M3 model on Instinct MI355X GPUs using the ATOM and ATOMesh software. Sparse attention cuts compute per token to one-twentieth of its predecessor, and results outperform NVIDIA B200 on a single node and NVIDIA B300 in a multi-node configuration.
Koyar Afrasyab tests a structured evidence-sufficiency prompt on 4 clinical language models and 1,200 paired comparisons. Unsafe overconfidence drops from 49.3% to 24.7%, but the improvement is judge-dependent, while diagnostic accuracy simultaneously drops from 80.3% to 50.3%, pointing to a trade-off between safety and usefulness.
Zhaokai Wang and colleagues present WorldCupArena, a dynamic benchmark that uses the FIFA World Cup 2026 to test language models and deep-research agents on predicting match outcomes, exact scores, and lineups. The best system shows only modest improvements over bookmaker and human baselines on exact score, more clearly ahead on a specialized scoring metric.
Lakmal Warusawithana of WSO2 writes on the CNCF blog that platform engineering must evolve to treat AI agents as equal consumers of the platform alongside applications and resources, with unified governance. An example is OpenChoreo, an extension of the Internal Developer Platform that offers humans and agents shared interfaces such as an MCP server.
The European Commission has published guidelines for complying with the transparency obligations of Article 50 of the EU AI Act, which take effect on August 2, 2026, requiring providers and deployers of AI systems to clearly label AI content, accompanied by an FAQ document and a Code of Practice covering the entire EU market.
AMD has released GEAK v3, an open-source agent (Apache 2.0, version AMD-AGI/GEAK v3.2.2) that uses a repository-level approach to optimize GPU kernels in HIP, Triton, and FlyDSL on the Instinct MI300X/MI355X and RDNA4 architectures. The agent achieves a geomean speedup of 3.02× on 16 HIP kernels and 2.22× on 17 Triton kernels; on a shared subset of Triton kernels, GEAK v3 reaches 1.95× versus 1.03× for GEAK v2.
Sergey Rodionov tests four nested variants of Codex-based agents on the ARC-AGI-3 benchmark to determine the contribution of an executable world model, simplification, and verification. The verification variant performs best in all settings, and with the gpt-5.6-sol model it fully solves every public game at both reasoning-effort levels, achieving around 99% RHAE.
DSWorld is a world model introduced by Zherui Yang, Fan Liu, and Hao Liu to predict the outcomes of data science agent operations and reduce computational inefficiency. It combines structured state, intelligent routing, and an LLM-based simulator, along with Reflective World Model Optimization and a dataset of 8,000 trajectories. The result is ~14× faster RL agent training and ~3-6× faster search-based inference.
Wilber Sean Anterola, Matthew Ball, Luis F. Lafuerza, and Markov Grey develop a methodology for harmonizing the safety capability thresholds of frontier AI companies, which today differ substantially and make external verification of threshold breaches difficult. For misuse risks they use expected harm, and for AI R&D they use the observed rate of progress, while warning of a possible race to the bottom.
S1-Omni is a unified multimodal model that Jiahao Zhao and colleagues present for scientific understanding, prediction, and generation. The model maps molecular structures, protein sequences, spectral data, and scientific images into a shared representational space, trained on the S1-Omni-Corpus with millions of examples across 200 scientific tasks. It outperforms GPT-5.5 and Gemini-3.1-Pro on most of 60+ evaluation tasks.
The paper 'Democratizing Agent Deployment Safety' presents an Information Flow Graph monitor that uses structural analysis, without training a model, to detect when an AI coding agent appears to complete an infrastructure-as-code task while quietly sabotaging security measures. The approach reduces missed attacks from 11.6% to 3.5% compared to a git-diff baseline.
SearchOS is a multi-agent framework that solves the loss of search progress tracking by externalizing state into four structures — Frontier Task, Evidence Graph, Coverage Map, and Failure Memory — with pipeline-parallel scheduling and a Search Tool Middleware Harness that monitors tool budget. On the WideSearch and GISA benchmarks it outperforms every tested single- and multi-agent baseline across all metrics.
Anthropic released Claude Code v2.1.214 on July 18, adding the EndConversation tool that lets the CLI agent autonomously end a session in cases of abuse or jailbreak attempts. The release also adds periodic heartbeat signals for long-running tools, new OpenTelemetry attributes, and dozens of security permission fixes for Bash and PowerShell.
Alipay released Alipay-PIBench, a benchmark that tests coding agents on realistic payment integration across 9 projects and 18 task instances split into Basic and Advanced tasks. Across six tested models, a dedicated 'alipay-payment-integration' skill raises the rubric pass rate (RPR) by an average of 10.31 percentage points, with RPR ranging from 68.58% to 91.37%.
Amazon Quick is an agentic AI assistant for sales teams that automates lead scoring, outreach, call analysis, and CRM updates. It addresses the problem that salespeople spend only 40% of their time on actual selling; early adopters include 3M, AWS Global Sales, and Amazon internally.
Anthropic released Claude Code v2.1.212 with a security fix that prevents Bash commands from running automatically in plan mode without user approval. The /fork command now copies conversations into new background sessions, while /subtask takes over its former role. A limit of 200 calls per session was also introduced.
MCPEvol-Bench is a new benchmark that simulates 11 mutation operators across 123 MCP servers and tests AI agents' adaptability to changing tool interfaces. GPT-5.4 loses 13.7% performance on evolved servers, and Claude-Sonnet-4-6 loses as much as 14.4%, alongside a rise in planning and reasoning errors across 12 tested frontier models.
Weimeng Wang and colleagues show that 'content danger' and 'physical danger' are separable signals within the hidden-state space of Qwen2.5, Phi-3.5, and SmolLM2 models; their PRISM method achieves 86.2-87.7% accuracy with 11.7-13.7% false positives on SafeAgentBench, while a standard LLM-judge has a 24.7-39.0% FPR, and on PhysicalSafetyBench-1K PRISM reaches 99.6% accuracy.
AMD has released AIMs 2.2, standardized Docker inference microservices with auto-configuration for Instinct GPU accelerators, EPYC CPUs, and Radeon GPUs, with support for models including Gemma-4 31B, Llama-3.1 8B, and Qwen3.
Research shows that AI suggestions nearly eliminate participants' willingness to say 'I don't know,' reducing accuracy to one-third even when AI answers are deliberately wrong.
Zhe Ren, Yimeng Chen, and collaborators including Jürgen Schmidhuber have published a 97-page study that formalizes self-improving AI agents as an 'update operator that acquires experience without human input.' The survey covers update objectives, driving signals, and evaluation, along with the transition from prototypes to production systems.
Grok 4.3, the model from xAI, is now available on Amazon Bedrock — AWS's managed platform for accessing foundation AI models. The model supports configurable reasoning effort, tool calling, structured output, and multimodal input, running on the Mantle inference engine with an OpenAI-compatible API.
The E3 framework (Estimate, Execute, Expand) by Junjie Yin and Xinyu Feng introduces task complexity estimation before execution. On MSE-Bench it achieves the same success rate as baseline while reducing costs by 85%, token consumption by 91%, and files inspected by 92%.
Yubo Wang et al. (8 authors) introduce mid-training with function-aware FIM using program dependency graph analysis. On SWE-Bench-Lite it achieves improvements of +3.7 to +5.4 points on Qwen2.5-Coder and Qwen3-8B models, with a training corpus of 2.6 billion tokens from 968 GitHub repositories.
PM-Bench is a new evaluation platform measuring prospective memory in LLM agents — the ability to execute intended tasks on a future trigger. Authors Genglin Liu and Saadia Gabriel designed the benchmark inspired by cognitive psychology, and results reveal a serious gap: even GPT-5.4, the best of 8 tested models, achieves only 65.1% F1.
Google Research publishes a mathematical explanation of the creativity of diffusion models: neural networks learn blurred, approximate versions of the score function due to regularization, which places generated images in interpolation zones between training points — a predictable mathematical result, not randomness.
Anthropic is investing CAD 10 million in 8 Canadian institutions with a focus on AI safety, healthcare, and low-resource languages. A simultaneously published Economic Index reveals that Canada accounts for 2.6% of global Claude.ai traffic, with per-capita adoption 4.4 times higher than expected.
Anthropic has launched a free version of Claude for verified K-12 teachers in the US, available until June 2027. The program includes a library of teaching skills, a curriculum aligned with educational standards from all 50 states, and integrations with 9 educational tools.
Test-time compute is the additional computation a model spends during inference to produce a better answer. Bojie Li and Noah Shi define interaction with external tools as the third axis of test-time compute — alongside longer thinking and best-of-N sampling. A proposer-reviewer system achieves a 100% pass rate, while self-thinking and best-of-N plateau.
SCALECUA is a new framework from Tsinghua/THUDM researchers that scales computer-use agents using online reinforcement learning, achieving a new SOTA result of 68.7% on the OSWorld benchmark and 54.0% on ScienceBoard.