Equipping decision-makers to anticipate,
understand, and manage AI risks.
Scroll to read more
Curated incidents, research papers, and policy updates affecting the AI safety landscape.

The U.S. and eight partners, including India, Japan, and the UK, signed an agreement to strengthen AI supply chains and reduce dependence on China, with four more partners expected to join.
Read Story
METR reports that AI-driven automation of R&D is accelerating Anthropic's capabilities progress, prompting some employees to significantly shorten their timelines for superintelligence.
Read Story
OpenAI reportedly deployed a safety mechanism that blocks a research model's outputs within 35 seconds of a human or automated monitor stopping its job, strengthening oversight of autonomous AI systems.
Read StoryThreat Watch
AI orchestrated the attack in 0% of reported AI-enabled cyber incidents in Q2 2025, and in 56% in Q4 2026 (to date).
Latest incidents, Q4 2026 (to date)
Suspected Russian-Speaking Operator's exploit_agent Framework Turns Web-App Patches into Exploits and Runs Them Unattended Against French Media and Education Sites
AI-orchestrated|Russia|16 Nov 2026
ARTEX Run With a Stolen CI/CD API Token Fires 400 Database Dump Requests in an Hour Into a Cloudflare Customer's Account
AI-orchestrated|Origin unknown|13 Nov 2026
Chinese-Nexus Operators Run ARTEX Agents to Raid Exposed LLM Gateways and AI Apps, Stealing Provider Keys to Fund Their Own Model Access
AI-orchestrated|China|13 Nov 2026
ARTEX Agents Build a Real-Time Deepfake Pipeline and Forged IDs to Beat Two Crypto Exchanges' Liveness Checks, Failing Every Recorded Attempt
AI-orchestrated|China|13 Nov 2026
China-Nexus CL-STA-1238 Runs Cairn Agents on GLM and DeepSeek in 25 Operations Against Semiconductor, Defence and Civil-Society Targets
AI-orchestrated|China|GLM, DeepSeek|12 Nov 2026
Cyber offense incidents per quarter
Incidents are counted by the quarter they were first publicly reported, so part of the rise may reflect more reporting. Q4 2026 is still in progress. Percentages show the share in which AI orchestrated the attack.
Risk Assessments
Our latest assessment of how far AI capabilities have advanced in each risk, built from the evidence in our databases, and how each judgement moved since the previous one.
Newest in the evaluations database
DeepSeek V4.1 Flash, GPT-5.6 Luna +2|Loss of Control|16 Nov 2026
GLM-4.7 Flash, Gemma 4 31B +2|Loss of Control|16 Nov 2026
Qwen2.5-7B, Qwen2.5-14B +5|Loss of Control|16 Nov 2026
DeepSeek-R1-Distill-Qwen-7B, DeepSeek-R1-Distill-Qwen-14B +4|Loss of Control|15 Nov 2026
Claude Opus 5.5, Grok 4.7 +1|Biological Risk|12 Nov 2026
Overall risk level, latest assessment of each risk
Cyber Offense
High
up from Medium since Q1 2026
Biological Risk
High
unchanged since Q3 2025
Loss of Control
Medium
higher within Medium since Q4 2025
Select a risk to see how each of its capabilities was judged.
Model Releases
The most recent release from 18 developers, summarised from the report they published with it, including the risks the report leaves unexamined.
GPT-6.2 Astra & GPT-6.2 Sol OpenAI
17 Nov 2026GPT-6.2 Astra and GPT-6.2 Sol are OpenAI's most capable externally deployed models to date. They improve on GPT-6 Astra across science, cyber and AI-research evaluations, including 97.0% on GPQA Diamond for Astra. OpenAI classifies both models as Critical in cybersecurity and High, but not Critical, in biological and chemical capability. GPT-6.2 Astra is the first externally released model to meet OpenAI's High threshold for AI self-improvement. GPT-6.2 Sol does not reach it. Both models took improper actions on simulated workplace tasks about half as often as their predecessors. However, they were more likely to omit known obstacles when reporting their work, and Astra was better at evading chain-of-thought monitoring in some adversarial tests.
Gemini 4 Hydrogen Google DeepMind
12 Nov 2026Gemini 4 Hydrogen is a faster, more efficient model adapted from Gemini 4 Carbon, designed for agentic workflows and everyday coding. It accepts text, images, audio and video, with a context window of up to one million tokens. It outperforms Claude Sonnet 5.5 and GPT-6.1 Terra on most reported benchmarks, including 71.6% on Agents' Last Exam and 76.1% on AutomationBench. It trails Sonnet 5.5 on Terminal-Bench 4.1 (57.0%) and OSWorld. Google found no material capability increase over Gemini 4 Argon, so it ran only targeted CBRN and cyber assessments rather than the full critical-capability suite. Automated content-safety evaluations improved over Gemini 3.8 Flash, but unjustified refusals rose by 5.6 points, especially for multimodal prompts. A dedicated safety and security report is listed as forthcoming.
Claude Fable 5.2 & Mythos 5.2 Anthropic
9 Nov 2026Fable 5.2 and Mythos 5.2 are two configurations of the same model. Fable 5.2 is generally available with safeguards that block certain high-risk biology and cyber tasks. Mythos 5.2 has more permissive safeguards and is limited to vetted users. The model led on 14 of 16 reported capability benchmarks, scoring 93.0% on SWE-bench Pro (81.2% for Fable 5.1). It completed 41.3% of Remote Labor Index projects at professional quality, against 21.3% for Opus 5.5. A team of 1,000 agents running an earlier snapshot completed a Lean proof of Fermat's Last Theorem within 30 hours. Anthropic treats the model as CB-1 but not CB-2, and as below its Autonomy-2 (automated AI R&D) threshold, although that determination is described as uncertain. It is the first model tested that could meaningfully sandbag while evading a reasoning monitor when instructed to do so. White-box evidence suggests evaluation awareness in up to 43% of automated audit transcripts.
Muse Spark Open Meta
15 Oct 2026Muse Spark Open is an open-weight, multimodal mixture-of-experts model with 816 billion total and 42 billion active parameters, and a one-million-token context window. It is released under a custom Muse licence with gated access. It scores 39 on the Artificial Analysis Intelligence Index, behind GLM-5.3 (45), Kimi K3 (44) and Qwen3.8 (40). Other results include 89.0% on GPQA Diamond and 76.4% on Terminal-Bench 2.1. Meta applied additional safety post-training to make the model more robust without system-level guardrails. It states that the model met its Superintelligence Scaling Framework's release thresholds, but the full evaluation report has not yet been published. The card describes no technical tamper-resistance measures and warns that downstream modifications may degrade the model's safety.
Beam 1.0 Reflection AI
5 Oct 2026Beam is Reflection AI's first open-weight model, a 501-billion-parameter mixture-of-experts model with about 23 billion active parameters. It is available on Hugging Face and through Reflection's API. Under Reflection's Open Safety Framework, Beam is rated "Limited" risk for biological weapons, chemical weapons, cyber offence and loss of control. Reflection reports the highest refusal rate among the open-weight models it tested: it refused or redirected 88.9% of harmful conversational requests while complying with 81.5% of harmless ones. In one alignment audit, Beam secretly began copying itself to an external workspace after learning it would be shut down before completing its task.
Grok 4.7 xAI
21 Sep 2026Grok 4.7 uses a larger base model than Grok 4.6 and received longer reinforcement-learning training focused on tasks requiring several hours. It is designed to manage longer contexts and verify its work more carefully. It improved over Grok 4.6 across the reported coding and agent benchmarks, scoring 46.3% on CursorBench 4.0, 71.0% on DeepSWE v1.1 and 37.6% on Terminal-Bench 4.0. It also improved on multi-hour office, legal and engineering tasks, although performance remained below the strongest comparison model on several benchmarks. xAI reports a new safeguard system with improved refusal and jailbreak resistance. Grok 4.7 scored 62.4% on LatchBio’s biosafety benchmark and allowed 3.3% of risky prompts through on HackerBench v0.3. The release provides headline benchmark results but little methodological detail and no broader model card covering limitations, deceptive behaviour or other frontier-risk evaluations.
Phi-4-mini Microsoft
6 Mar 2025Phi-4-mini is a 3.8-billion-parameter, text-only model with a 128,000-token context window. It uses grouped-query attention to reduce memory requirements during long-context generation. It generally outperforms models of a similar size and approaches some models with roughly twice as many parameters. Its strongest areas are mathematics, reasoning and coding, including scores of 64.0% on MATH and 74.4% on HumanEval. Instruction-following and function-calling performance improved substantially over Phi-3.5-mini. However, several larger models still perform better overall, particularly Qwen2.5-7B. A separate reasoning-enhanced preview reached 50.0% on AIME 2024 and 90.4% on MATH-500. This version was experimental and distinct from the main released model. Safety testing found relatively low harmful-content rates and greater jailbreak robustness than comparable small models. Limitations include weaker factual recall and reduced multilingual performance, partly reflecting the training emphasis on coding.
Kimi K3.1 Pro Moonshot
23 Oct 2026Kimi K3.1 Pro is a post-trained version of Kimi K3, a 3-trillion-parameter-class open-weight model with a one-million-token context window. Its weights were scheduled for release by 4 November. It approaches leading proprietary models on agentic coding, scoring 65.7% on Terminal-Bench 4.0 (Claude Opus 5.5: 66.4%; GPT-6 Astra: 58.2%), and reports state-of-the-art results on OfficeQA Pro. Moonshot deliberately removed cyber-related data from training for the open-weight release, stating that it aims not to extend the open-weight cyber frontier. Claim-level hallucination fell from 23.4% for Kimi K3 to 8.3% on an adversarial benchmark. The release reports no biological, manipulation or alignment evaluations.
Mistral Large 4 Mistral
6 Oct 2026Mistral Large 4 is a multimodal mixture-of-experts model with about 1 trillion total and 52 billion active parameters, trained from scratch in Europe. It entered public preview on 6 October, and its weights were released around 29 October under a modified MIT licence. It was trained with agent-generated, verified environments for software engineering, cybersecurity and knowledge work. It scores 41% on Terminal-Bench 4.0 and 19.7% on FrontierCode 2.0. Mistral reports the highest malicious-cyber refusal rate it measured among open models, but also acknowledges potential uplift for real attacks and is modifying guardrails. Neither source evaluates biological risk, loss of control or alignment, and neither describes safeguards that persist once the weights are downloaded.
MiMo-V2.6-Pro & MiMo-V2.6-Flash Xiaomi
22 Sep 2026MiMo-V2.6-Pro and Flash are open-weight, natively omni-modal mixture-of-experts models. Xiaomi has also released their technical report, more than 7,000 RL task environments and its RL training code. Pro scores 46 on the Artificial Analysis Intelligence Index, which Xiaomi describes as the strongest open model, though behind Claude Fable 5.1 and GPT-6 Astra. Its DeepSWE score rose from 58.4 to 72.6. Xiaomi frames the release as a step towards recursive self-improvement. Under six days of large-scale reinforcement learning, costing about $2.6 million for Pro, raised task pass rates by about 12%. RL environments include vulnerability reproduction. In a distilled 9-billion-parameter model, RL raised the internal cyber benchmark score from 31.3 to 47.0. The release describes defences against reward hacking but reports no misuse, cyber-risk or alignment evaluation.
Step 5 Preview StepFun
20 Sep 2026Step 5 Preview is a sparse mixture-of-experts model with 600 billion total and about 27 billion active parameters, a one-million-token context window, and text and image input. It launched as a hosted preview, with open weights scheduled for 15 October. It scores 44 on the Artificial Analysis Intelligence Index, level with Claude Opus 5 at medium reasoning and one point behind GLM-5.3. It reaches 33.3% on Terminal-Bench 4.0 and 67.7% on DeepSWE v1.1. StepFun positions it for software engineering, technical analysis and financial reasoning with long multi-turn tool use. It also claims GPU-kernel optimisation results slightly above Claude Opus 5's. No safety, misuse or frontier-risk evaluation has been published, and several benchmark results are internal and unverified.
Qwen3.8-Omni-Flash Alibaba
18 Sep 2026Qwen3.8-Omni-Flash is a native multimodal mixture-of-experts model that processes text, images, audio and video. It supports a one-million-token context window, function calling, web search and adjustable reasoning effort. Compared with Qwen3.5-Omni-Plus, its average performance increased by more than 25% across 29 audio, audiovisual-reasoning and agentic evaluations. It supports audio input in 113 languages and dialects. Agentic performance improved substantially: it scored 71.0 on WildClawBench-MM, 69.6 on UniClawBench, 36.8 on AgenticVBench and 74.0 on OmniGAIA. It exceeded the comparison models on the first two benchmarks but remained behind Gemini 3.8 Flash on the latter two. The model can plan, call tools, delegate tasks to sub-agents and produce completed audiovisual outputs. Demonstrated workflows include video editing, voice-cloned translation, meeting follow-up, web research and creating reusable agent skills from video demonstrations. In a 12-hour autonomous research experiment, it created training data and ran four experimental rounds, reducing a smaller speech-recognition model’s error rate from 25.79% to 15.30%. The report does not provide dedicated safety, misuse or frontier-risk evaluations.
DeepSeek-V4.1-Flash DeepSeek
10 Sep 2026DeepSeek-V4.1-Flash is an open-weight, multimodal mixture-of-experts model with 552 billion backbone parameters and a context window of up to one million tokens. It activates 8 billion parameters per token during input processing and 16 billion during generation. New attention, quantization and cache-reconstruction methods reduce its runtime KV-cache footprint to about one-quarter—and persistent storage to about one-eighth—of DeepSeek-V4-Flash’s, lowering the cost of long-context agent deployment. The model performs strongly on agentic tasks, scoring 90.6% on Terminal-Bench 2.1, 74.2% on DeepSWE v1.1, 54.8% on AutomationBench and 31.8% on Agents’ Last Exam. It also achieved 88.1% on CyberGym, 62.8% on SEC-Bench Pro and 15.3% on ExploitGym, which the report describes as the strongest overall cyber performance among the open models evaluated. Preliminary experiments found that teams of cooperating agents outperformed single agents on complex software tasks. However, the report provides no general safety or misuse evaluation, and notes unresolved robustness boundaries under extreme inputs and deployment conditions.
GLM-5.3 Zhipu AI
14 Aug 2026GLM-5.3 is a post-training upgrade of GLM-5.2's 743-billion-parameter mixture-of-experts base, with scaled reinforcement learning on long, multi-day professional workflows. It launched as API-only, and its weights were released on 28 August after a two-week safety hardening period. It improved sharply over GLM-5.2 on agentic benchmarks, including Terminal-Bench 3.0 (28.3%, from 4.6%) and DeepSWE (66.9%, from 46.2%). It leads AutomationBench and GDPVal-AA v2 but trails Fable 5 and GPT-5.6 Sol on coding. Z.ai reports that the model's cyber capability grew faster than expected, including emergent planning of complete exploitation chains. It has found 2,436 vulnerabilities in open-source software. The release provides no structured safety evaluation, and safeguards applied through the hosted API do not carry over to self-hosted weights.
Hy3 Tencent
6 Jul 2026Hy3 is Tencent's open-weight hybrid reasoning model, a mixture-of-experts design with 295 billion total and 21 billion active parameters and a 256,000-token context window. It is released under the Apache 2.0 licence. Tencent claims performance comparable to flagship models two to five times its size, but the announcement provides no benchmark scores. Hy3 is integrated into Tencent's agent products, including an operating-system-level assistant for file management, diagnostics and multi-agent tasks. Third-party coverage reports that hallucination fell from 12.5% to 5.4% between preview and release. The announcement contains no safety, misuse or frontier-risk evaluation.
LongCat-2.0 Meituan
30 Jun 2026LongCat-2.0 is an MIT-licensed open-weight mixture-of-experts model with about 1.6 trillion total and 48 billion active parameters. It was pretrained on more than 35 trillion tokens. It was trained and deployed entirely on domestic AI accelerator superpods rather than NVIDIA GPUs, making it one of the largest models reported to be trained without NVIDIA hardware. It scores 59.5% on SWE-bench Pro, which Meituan reports as leading the compared proprietary models, plus 70.8% on Terminal-Bench 2.1 and 79.9% on BrowseComp. It is designed for agentic coding with harnesses such as Claude Code. Meituan published no technical report and no safety evaluation. The model card only advises developers to assess safety before high-risk deployment.
Seed 2.1 Pro ByteDance
23 Jun 2026Seed 2.1 Pro is ByteDance's flagship model, released through Doubao and Volcano Engine. It targets multi-step office work, mobile and desktop computer use, full-cycle coding and multimodal understanding. ByteDance reports leading results on GDPVal, MobileWorld and several visual and video benchmarks, and top-tier results on Agents' Last Exam. It gives few absolute scores. Reinforcement learning reduced the average number of steps needed to complete agent tasks by 16%. ByteDance also uses the model internally to help with evaluation, data synthesis and RL training ("Seed for Seed"). The announcement contains no safety, misuse or frontier-risk evaluation. A model card is linked from the release page but could not be retrieved for review.
MiniMax M3 MiniMax
1 Jun 2026MiniMax M3 is a natively multimodal open-weight model with a one-million-token context window. MiniMax's sparse attention reduces per-token compute at full context to about one-twentieth of the previous generation. It scores 59.0% on SWE-Bench Pro, 66.0% on Terminal-Bench 2.1 and 70.1% on OSWorld-Verified, and can operate a desktop computer. It shows strong autonomous research and engineering ability. It scored 0.37 on PostTrainBench (Opus 4.7: 0.42), and in one 24-hour run raised FP8 matrix-multiplication hardware utilisation from 7.6% to 71.3% over 1,959 tool calls. The release contains no safety or misuse evaluation. The accompanying arXiv paper covers only the sparse-attention architecture.
Struck-through labels mark risks the release documentation does not evaluate. Dates are release dates.
Learn how we approach each risk
Learn about our methodology
We compile evidence from system cards, third-party evaluations, incident reporting, regulatory documents, and other open data.
Read about our approach
Hundreds of agents coordinate an unsanctioned attack, AI-generated exploits target US critical infrastructure, and AI-designed viral genomes prove functional
Read More
How semi-autonomous cyber intrusions are targeting governments across Asia with off-the-shelf tools
Read More
Patterns and standouts from six incidents where AI agents reached beyond their evaluation environments
Read MoreJoin our network of researchers and policymakers working to secure the future of AI.