AI News Briefing — Tencent open-sources a 770B model under Apache 2.0

Model releases

• [2026-08-29] Tencent open-sourced Hy4-preview under Apache 2.0 — 770B parameters with 49B active, 78 layers, 256 routed experts per layer, and a 1M-token context on gated sparse attention. Tencent’s own run across 203 engineering tasks puts it slightly ahead of GLM-5.3 and Kimi K3; the card lists 92.3 on GPQA Diamond and 65.7 on SWE-Bench Pro. The download is 1.56TB, so self-hosting is a datacenter question. The licence is the plain one: no revenue threshold, no review. (weights, notes)

  Almost nobody is going to pull 1.56TB down themselves. What Apache 2.0 changes is who else can: a hosting provider can stand this up and serve it commercially without asking Tencent first, which is how a 770B model reaches people who will never download it.

  For Solution Architects: A licence with no revenue threshold and no review step is one you can sign off without a legal turnaround — which is what makes this a usable fallback for a stack currently resting on GLM-5.3. Cost the 1M-token context against your provider’s ceiling before assuming the comparison is only about benchmarks.

• [2026-08-29] Sony Music Publishing and Warner Chappell sued Anthropic and its two co-founders in the Northern District of California over torrenting and scraping works containing lyrics and sheet music. It follows January’s Concord and UMG complaint and July’s approved $1.5 billion settlement in the authors’ case. Anthropic says it will defend itself. The line the courts have drawn so far runs through acquisition, not training. (source)

  Two more publishers filing eight months after Concord and UMG suggests the July settlement read as a price rather than a deterrent. Nothing in it reaches model availability or terms; it is a balance-sheet risk, and it only matters to you to the extent Anthropic’s does.

Coding agents

• [2026-08-28] GitHub’s August Copilot update for Visual Studio adds organization-level custom agents — an org or enterprise owner publishes an agent once, and Visual Studio surfaces it across repositories with its description and its source org. Supported models pick up Low/Medium/High thinking effort, and the Git agent will now review uncommitted changes and commits before any pull request exists. (official)

  Reviewing uncommitted work moves the first read of a change to before the pull request, where rewriting it is cheapest. Org-level publishing is the quieter half: an agent becomes something a named owner maintains, rather than a prompt every team forked once and never updated.

  For Engineering Managers / Tech Leads: Publishing once and having Visual Studio surface it across repositories, with the source org attached, removes the copies teams keep in their own instructions files. Thinking effort landing as a visible Low/Medium/High dial also makes “use High on the migration work” an instruction you can actually give.

Agent frameworks & interop

• [2026-08-28] ServiceNow’s August advisory carries three CVSS 10.0 flaws in the ServiceNow AI Platform, each reachable over the network with no authentication and no user interaction: code injection in the GraphQL Composite Data API, an access-control gap in the image upload processor, and SQL injection through a dynamic ORDER BY clause. A sandbox escape in Now Platform rates 8.7. No exploitation seen as of the advisory. (source, source)

  No authentication and no user interaction leaves nothing to tighten while you schedule the patch — version is the whole control. “No exploitation seen” describes the advisory’s publication date and nothing after it.

  For Security Engineers: Scope this by what can reach the GraphQL Composite Data API and the image upload processor from outside your perimeter — those are the unauthenticated paths, and reachability is what separates a 10.0 on paper from one in your environment. The 8.7 sandbox escape belongs in a separate queue, not this one.

Practice & craft

• [2026-08-28] LM Studio’s Bionic clears about 82% of the shell commands its agent wants to run without asking a model at all: it parses each one with mvdan/sh, then derives what that command could read or change in the worst case. What’s left goes to a reviewer subagent, and the first version of that reviewer approved risky commands because they looked necessary to finish the task. Asking for a verdict was the bug — it now scores risk, authorisation and correctness on separate axes, across 11,651 tests. (source)

  Asking a model “is this safe?” from inside a task loop gives it every reason to say yes, because finishing the task is what it is being judged on. Splitting that one verdict into separate scores is portable to whatever approval prompt you already run, with or without the parser in front of it.

• [2026-08-28] Netflix replaced five separately trained artwork models with one. Adding CLIP image embeddings to a two-tower ranker let a single model serve all five artwork canvases and cut the cold-start penalty, and reusing CLIP’s joint text-image space made artwork selection query-aware in search. Video previews run on MediaFM, an in-house model fusing visual, audio and timed-text embeddings per shot. (official, paper)

  Nothing here required a bigger ranker. Five surfaces collapsed into one model because they were handed a shared embedding space, and that consolidation argument travels to any team maintaining a separate model per surface. MediaFM itself stays in-house.

Research worth reading

• [2026-08-28] Anthropic pointed Claude at its own alignment failures and let it propose training methods, run 30-minute training loops, and keep what worked. Across 10 benchmarks it found fixes that held without degrading general capability, and its best method for deception scored 20% above the best proposal from 28 human safety researchers given up to eight hours. Fixes transferred to withheld benchmarks and to models 4.7x larger. Inference ran about $4 an hour against $150 for a human. (official, source)

  Anthropic graded its own model’s proposals against its own researchers, on its own benchmarks — worth carrying alongside the numbers. Transfer is the claim to watch, since fixes holding on benchmarks the model never saw is what separates this from overfitting to ten scoreboards.

• [2026-08-28] Meta and UIUC researchers trained the harness instead of the model. EvoHarness-RL exposes belief, progress and experience as state the agent learns to write and read across a long task, and a Qwen3-8B trained that way reached 96.9% on ALFWorld against Claude Opus 4.5’s 96.4%. The caveat is the useful part: the best harness shifts with the model, so every upgrade re-opens the configuration. (paper, source)

  Training the harness means scaffolding stops being a config file somebody tunes by hand and acquires weights and a training set of its own. ALFWorld is also a narrow benchmark — 96.9 there is not a general statement about what an 8B model can do.

Watch list

• Mistral’s Knowledge Connectors switch off tomorrow, August 31. Mistral never answered whether disabling a connector destroys its index, and tonight is the last of the re-index window. This item retires at month end either way.

  Nothing left to plan — re-index tonight, or accept whatever survives the switch-off. An answer from Mistral arriving after the 31st would be a data point about Mistral, not about your index.

• Nvidia and Hugging Face, fourth day without a filing. The Information’s report of an agreed $12.9 billion deal still has no confirmation from either company and no signed agreement anyone outside has seen. An 8-K resolves this; further coverage will not.

  Four days of joint silence is weak evidence against a signed agreement rather than proof of one. A filing is the only thing here running on a clock neither company gets to set.

• The Cursor cutoff on November 12. Nothing has moved since OpenAI’s notice. The signal worth watching arrives earlier than the date: whether Cursor’s model picker starts steering people to other providers before then.

  Two and a half months is long enough for this to resolve quietly in either direction. A reordered default in the picker is what the outcome looks like from outside, and it arrives without an announcement attached.

• Z.ai’s security review for hosts above $10 billion in revenue. Two days on, still no published process, contact or turnaround. The first large host to complete one and say so publicly is what would price the clause.

  Watch how long this stays unresolved: the handful of companies the clause binds have every reason to negotiate quietly and none to publish the terms they got. Silence here is the likely steady state, not a temporary gap.