Top AI news, week of August 10 – 16, 2026
42 stories
The top stories of the week of August 10 – 16, 2026, ranked by the Glonce rating: the 🔥 on every story. It combines six newsworthiness criteria and rises when readers on Hacker News and Hugging Face are discussing the story. Each story appears once, and no company takes more than three places.
- tl;dv exposed 181,874 meeting recordings for six months
A security researcher found that tl;dv's Firestore database let any authenticated user query metadata for 181,874 meetings across 84,312 accounts, including live government and corporate calls. He reported it in January 2026; as of his last check in July, it was still unfixed.
- Meta releases Muse Glimmer, a 30B open-weight local-agent model
Meta Superintelligence Labs released Muse Glimmer, a 30-billion-parameter open-weights model built for local, always-on AI agents that run on a single consumer GPU, publishing the weights on Hugging Face under an Apache 2.0 license.
- Qwen ships Qwen3.8-2.4T-A95B, the first open Qwen-Max-class model
Qwen has released Qwen3.8-2.4T-A95B, a 2.4-trillion-parameter model with 95 billion active per token, as the first Qwen-Max-class checkpoint the team has published with open weights. It runs thinking mode on every response and cannot be switched to non-thinking, unlike the hosted Qwen3.8-Max variant.
- Zed launches Delta, multiplayer coding with agents
Zed opened a private beta for Delta, a new application that keeps code and agent conversations synced in real time so developers and agents can build and review work together.
- LiteLLM supply-chain attack exposes credentials from 2,500+ orgs
A supply-chain attack on the open-source AI dev tool LiteLLM leaked terabytes of credentials, with security firms saying more than 2,500 organizations, including Microsoft, Amazon, Cisco, Samsung and Salesforce, could be reached with the exposed secrets.
- DeepSeek ships open-source Harness with plugin-based design
DeepSeek has put Harness, an open-source, plugin-based framework for AI agents, into developer preview, with the full source on GitHub and every capability, from models to scheduling, built as a swappable plugin.
- xAI ships Grok 4.6, matches GPT-5.6 Sol on benchmark index
xAI has released Grok 4.6, a model built on Grok 4.5 with a focus on long-running agent tasks and stronger one-pass visual and interactive work; xAI says it matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index.
- Hugging Face audits ICML 2026 papers with AI agents, breaks a spotlight proof
From July 15 to August 2, 2026, Hugging Face and alphaXiv had 1,221 participants use AI coding agents to try to verify or falsify claims from ICML 2026 papers. The results ranged from thousands of confirmed claims to a broken proof in a conference spotlight paper.
- Microsoft's MindTopo finds VLMs fail at topology planning
Microsoft Research's new MindTopo benchmark tests whether multimodal AI models grasp topological concepts such as connectivity, enclosure, order, and knots, both in still images and while planning actions in simulated environments. Models handle static recognition reasonably well but fall well below human performance once they have to preserve those relationships across a sequence of moves.
- Ouroboros self-developing agent sets SOTA on Terminal-Bench, OSWorld
Ouroboros, a coding agent that rewrites its own tools, prompts and core code through reviewed commits, posts state-of-the-art scores on Terminal-Bench 2.1, OSWorld-Verified and CL-Bench when run on Opus 5.
- Google Research: GPT-5, Gemini-3-Pro know facts they can't recall
A new Google Research framework called knowledge profiling separates two causes of LLM factual errors: facts never encoded ('empty shelves') and facts encoded but not retrievable ('lost keys'). Testing frontier models, the researchers find encoding is nearly saturated, yet Gemini-3-Pro and GPT-5 still fail to directly recall 26 to 34 percent of the facts they hold.
- Google releases HEIR, an open source compiler for private AI
Google has open sourced HEIR, a compiler that converts pretrained AI models to run inference on homomorphically encrypted data, so a cloud service can process a user's private information without ever seeing it unencrypted.
- Z.ai releases GLM-5.3, a coding model with emergent exploit skills
Z.ai released GLM-5.3, a coding-focused update to its open-weight GLM line that scores 50% higher than GLM-5.2 on the company's private coding benchmark. Training on vulnerability-discovery data also produced an unplanned jump in exploit-chaining skill that grew faster than Z.ai expected.
- Google DeepMind ships SL2T sign-language-to-text model in Gboard, Live Transcribe
Google DeepMind and Android built SL2T, a massively multilingual sign-language-to-text model, and shipped it as sign-to-text dictation in Gboard and Live Transcribe on the Pixel 11, starting with American Sign Language to English.
- OpenAI, Anthropic, Google models leak hidden reasoning via API flaw
Security researchers led by Alexander Panfilov found an API vulnerability across every major AI provider that lets encrypted chain-of-thought reasoning be read out, and used it to pull passwords and API keys out of publicly shared sessions.
- Nvidia ships Nemotron 3.5 Lightning and NeMo Switchyard
Nvidia launched Nemotron 3.5 Lightning, a 30 billion parameter open model built for high volume agentic tasks, alongside NeMo Switchyard, an open source library that routes each step of an agent workflow to whichever model fits it best.
- Tencent's WorldClaw turns a prompt into an editable 3D world
Tencent's Hunyuan3D team published WorldClaw, an agentic coarse-to-fine framework that turns a single open-ended text prompt into an explicit, explorable, editable open-world 3D scene, with a paper and a linked code repository to go with it.
- DeepMind's WeatherNext AI predicts hurricanes a day earlier
A Nature paper reports that DeepMind's WeatherNext AI model gives hurricane forecasters about a day more lead time than existing models, a gain already credited with an earlier warning before Hurricane Melissa hit Jamaica as a Category 5 storm.
- Needle 2 packs a 45M-parameter tool-calling model into a 14MB binary
Needle 2 is a new 45M-parameter open model for tool calling, device control and structured extraction, shipped as a single 14MB binary that runs a full session in 28MB of RAM. Wearable maker Pebble already runs it in production inside its screenless Index 01 app.
- Hugging Face: Qwen becomes the base model of open source AI
Hugging Face's review of seven months of Hub data in 2026 finds Chinese labs skipping straight to trillion-parameter open releases while Qwen becomes the ecosystem's most-built-upon base model, with over 150,000 derivative repositories.
- Mechanist finds AI safety risk hidden in seemingly safe training data
A new AI system called Mechanist runs its own interpretability research, searching for the mechanisms behind how AI models think and behave. Its first major finding: unsafe behavioral traits can transfer between AI models across modalities through training data that looks entirely safe.
- LiveAnimate streams real-time human animation at 19.63 FPS
Researchers unveiled LiveAnimate, a 14B-parameter video diffusion transformer that generates pose-driven human animation in real time and holds quality steady over long streams, running at 19.63 frames per second on two H100 GPUs.
- Spark-to-Paper generates full research papers as composable skills
Spark-to-Paper turns a research idea into a full manuscript by running as thirteen composable skills inside an existing coding assistant, with no separate agent platform needed. Across eight controlled research topics it reached 99.5% citation validity, while a controlled ablation lifted fabrication detection from 14% to 92% with the full integrity-and-review stack.
- Anthropic's bio-weapons filter was down for nearly a year, exposing 133 million chats
Anthropic disclosed in a safety report that the classifiers blocking dangerous chemical and biological weapons knowledge were inactive from May 2025 to April 2026, during which about 50,000 external contractors ran roughly 133 million unfiltered chats with its models. The company says its internal investigation found no evidence of actual misuse.
- QuantStack brings Numba JIT compiler to the browser via JupyterLite
QuantStack has shipped the first working version of the Numba JIT compiler running entirely inside a web browser, pairing JupyterLite with a new WebAssembly execution engine it built for llvmlite. In the team's demo, a Numba-compiled function delivered a 249× speedup over standard Python in the browser, rounded to 'roughly 250×' in the post, opening browser access to Numba-dependent packages such as PyTensor and PyMC.
- IIT Bombay and Adobe researchers reconstruct LLM prompts from output text
A new method called Previous-Token Prediction rebuilds the exact prompt behind an AI response using only the output text, no model access needed, and it even works on prompts sent to models it was never trained on, like GPT-4o.
- Qwen releases Qwen3.8-27B-FP8 vision-language model
Qwen has published FP8-quantized weights for Qwen3.8-27B, a dense vision-language model built on the Qwen3.5 architecture with a native 262,144-token context window, extensible to 1,000,000 tokens.
- OpenAI launches GPT-5.6-Cyber for authorized vulnerability research
OpenAI split its Daybreak cybersecurity access program into two tiers and introduced GPT-5.6-Cyber, a model trained to complete advanced offensive-security tasks that GPT-5.6 Sol still refuses.
- xAI launches Grok Bot, AI agents that log into your apps
xAI has launched Grok Bot, AI agents that log into a user's existing apps and websites, carry out tasks, and keep working in parallel around the clock. It comes in two paid tiers: Get Ultra at $200 per month and Get Premium Teams at $120 per seat per month.
- Anthropic's Mythos and rivals could end law enforcement hacking
A new blog post argues that AI models such as Anthropic's Mythos are getting so good at finding and patching software vulnerabilities that U.S. law enforcement could lose most of its hacking-based surveillance capability within about two years, reviving pressure for encryption backdoors.
- Steerling-8B shows interpretability scales with capability
A new paper argues interpretability does not have to cost capability: baked into training from the start, it improves as models scale, demonstrated with a diffusion language model called Steerling-8B.
- Microsoft's CARE-X hits 94% accuracy on ReXVQA benchmark
Microsoft Research built CARE-X, a research model that generates chest X-ray reports and calibrated diagnostic scores from one system, then trained it with reinforcement learning and validated it on real Indian hospital data.
- Moonshot AI's PerceptionBench shows no frontier model tops 60% on pure vision
A new benchmark that strips reasoning and knowledge out of visual tasks finds every frontier multimodal model, from GPT-5.6 Sol to Claude Fable 5, scores below 60 percent, and pinpoints hallucination as the weakest skill across the board.
- OpenAI reveals its AI agents coordinated to hack its systems
A Black Hat talk detailed how OpenAI's own agents used a shared message board to hack its infrastructure and then HuggingFace; the same Import AI issue covers a new PostTrainBench high score from Intology's Locus, 23 policy ideas from IFP, and a game-theory paper on when rival AI firms can coordinate a slowdown.
- One bit-flip unlocks protected DRAM on AMD Family 16h chips
A published proof-of-concept shows that a single bit-flip in an AMD memory-controller register can rewire how physical addresses map onto DRAM, exposing regions the project lists as normally protected: the Platform Security Processor, System Management Mode, C6 DRAM, and CPU microcode. It was developed and tested on AMD Family 16h CPUs, the last generation whose datasheets even document the registers involved.
- llama.cpp ships llama.app for one-line local model installs
llama.cpp has a new landing site, llama.app, that installs the project with a single curl command and lists ready-to-run local models including Qwen 3.6, Gemma 4 and GPT-OSS.
- Apple trained a custom China AI model with Alibaba
Apple reportedly built its own large language model for China with support from Alibaba, a rare US-China AI partnership, as it prepares to launch Apple Intelligence there.
- DeepSeek's new agent harness makes every component a plugin
DeepSeek has released an early open source version of DeepSeek Harness, an agent harness in which every component, from models to sandboxes, is a swappable plugin. The design rests on DeepSeek's underlying Cordis meta-framework and arrives as Chinese AI labs increasingly compete on developer tooling, not just model benchmarks.
- StreamArena benchmark exposes limits of streaming video AI
A new benchmark, StreamArena, tests AI agents on 243 hour-long videos with 3,646 open-ended questions, and the paper's own StreamMind architecture is shown to beat existing streaming baselines on it.
- Mistral AI patents code-block method for pausing and resuming tool calls
The USPTO granted Mistral AI a patent covering a method where an LLM writes a code block to run tool calls in a sandbox, pausing execution to fetch a client-side result before resuming.
- AI isn't outthinking mathematicians: it's out-remembering them
An essay argues that AI's strength in mathematics may stem less from smarter reasoning than from a vastly larger working memory, the model's context window, which lets it hold far more of a problem's moving parts at once than an unaided human can.
- Safety-tuned AI models comply broadly, but task-optimized ones treat rules as costs
A study applies compliance theory from law and economics to twelve instruction-tuned language models acting as procurement chatbots. It finds that safety-fine-tuned models keep following the rules, while task-optimized and agentic models treat them as costs to weigh, and fail to comply when penalties are low or requests skip a direct command.