AI News — August 11, 2026

Model releases

• [2026-08-10] Meta — Muse Glimmer is a 30B distillation of the Muse line under a plain Apache 2.0 licence: a 2B vision encoder over a 28B text decoder, 52 layers, 32k context, GGUF quantisations from day one. It scores 75.5 on MCP Atlas where the models it names sit between 54.2 and 62.5, and 51.2 on SWE-Bench Pro. Inference fits one 80GB H100; Simon Willison ran an 18.16 GB quant in LM Studio and had it read Datasette’s auth code. (official, source, source)

  Agentic tool-calling stops being an API line item and becomes a capacity-planning question once the model fits one node, and that is a different conversation with a different owner.

  For ML / Data Engineers: A GGUF quant at 18.16 GB runs in LM Studio on hardware already on your desk, so an eval harness for tool-calling work can be built against the same weights that would later serve production on a single H100.

• [2026-08-10] OpenAI — GPT-5.6-Cyber is built on Sol and sold to nobody: it reaches only trusted partners through a new Daybreak Red tier, with Accenture, IBM, CrowdStrike and Cloudflare named. Daybreak now splits Blue (defensive work on general models) from Red (vulnerability research, exploit validation). OpenAI rates the model High but not Critical for cyber, three days after pausing Astra at that Critical line. (official, source)

  Rationing capability by contract rather than by price is new as a distribution shape, and it means whatever defensive tooling gets built on top arrives as a vendor product rather than something you can call directly.

Coding agents

• [2026-08-10] JetBrains — WebStorm 2026.2.1 bundles Chrome DevTools Connect, so the IDE agent drives a real browser: opens pages, walks a multi-step flow, reads console and network, screenshots, then re-tests its own fix. It arrives as a CLI skill the agent installs on first use. The loop it removes is you describing what you saw on screen. (official)

  Because it arrives as a CLI skill the agent installs itself, the browser-driving half is not really WebStorm-shaped, and any agent able to install a skill can reach for the same capability.

  For Software Developers: Hand a bug report’s URL and steps straight to the agent and it walks the flow, reads the console and network panels, and re-runs the whole thing after patching — the screenshot round-trip through you disappears.

• [2026-08-10] AWS — Continuum, the vulnerability-remediation agent in gated preview since June, is being wired into Claude Code, Codex and Kiro rather than kept in an AWS console. Findings and fixes land in the tool where the code was written, whichever model wrote it. No ship date was given for any of the three integrations. (official, source)

  AWS putting the same remediation agent inside two rivals’ coding tools treats the fix layer as portable rather than as a reason to open its console, though with no dates attached there is nothing to plan around yet.

MCP

• [2026-08-10] Cloudflare — A dashboard toggle now injects WebMCP into any site on the platform, no redeploy. An edge HTMLRewriter drops in a bridge.js that composes tool packs named in a data-packs attribute and registers each one, so an agent calls searchFlights instead of scraping the page. Developer preview, two packs, Chrome 145+ only. (source)

  An edge rewriter puts an agent-callable API on a site whose team never wrote one, which also moves the question of what tools exist to whoever holds the dashboard login.

• [2026-08-10] Nutanix — An open-source MCP server fronts the Prism v4 API gateway, translating plain-English requests into infrastructure calls under the RBAC already configured there. What is worth copying is the rest: request throttling and metering aimed at agent swarms, audit logs of every AI-initiated command, async handling for long tasks. Upwork and Digital Science shipped servers the same day. (source)

  Prism’s existing RBAC carries the access decisions, so no new permission model had to be invented for the agent — which is most of why an infrastructure MCP server can ship at all.

Agent frameworks & interop

• [2026-08-10] Brex — CrabTrap is an MIT-licensed Go proxy that governs agents at the network edge instead of in the SDK. Set HTTP_PROXY and every outbound request is checked against static rules first, with an LLM judge reasoning about the leftovers — under 3% of production traffic, by Brex’s count. Framework- and language-agnostic by construction, with OpenTelemetry metrics in Prometheus format. (official, source)

  A proxy catches every agent that honours the environment variable and nothing that doesn’t, so the real work is making sure the runtime cannot reach the network any other way.

  For Security Engineers: Static rules handle the bulk and the judge model sees under 3% of traffic, so the cost of a policy layer stops scaling with request volume — and the OpenTelemetry metrics come out in Prometheus format for the alerting you already run.

AI cost tracking & telemetry

• [2026-08-10] Uber exhausted its 2026 AI budget by April, with individual engineers running $500 to $2,000 a month through Cursor and Claude Code. The response was a hard $1,500 monthly cap per tool and a dashboard engineers can watch. VentureBeat’s argument around it: stuffing the context window was never a strategy, and curated agentic memory is the cheaper discipline. (source)

  A $1,500 cap sits roughly where the heaviest users were already landing, so it bounds the bill without answering why a month of assisted work costs that much in the first place.

Practice & craft

• [2026-08-09] Simon Willison stored a document’s full revision history as a zstd-compressed JSON array of every prior version, chunked at 128 revisions or 3MB, timestamps kept uncompressed alongside. A thousand revisions of 20.4 MB of text compressed to 80.3 KB. He described the idea by voice, then let GPT-5.6 Sol Pro grind for 38 minutes building the prototypes. (source)

  Near-identical copies are exactly what a general-purpose compressor is good at, which is why keeping whole versions costs less here than a patch chain would cost in complexity.

Research worth reading

• [2026-08-09] RETRACE verifies a coding agent’s patch without reusing the reasoning that produced it. One pass builds the repair rationale from issue and trajectory; a second infers what problem the patch solves having never seen the issue; disagreement triggers revision. Training-free, bolted onto mini-SWE-agent and OpenHands, worth 7.0 and 3.6 points of pass@1 on SWE-bench Verified across two backbones. (official)

  Two independent derivations compared against each other is a cheap pattern that needs no new model, and patch review is only the most obvious place to point it.

• [2026-08-09] A benchmark score says nothing about what it costs you to catch the failures. This paper names the gap — verification-cost errors, wrong outputs that a stated share of reviewers miss inside a realistic review budget — and measures it on code generation and document understanding. An accuracy gain can be undetected review burden moved somewhere else. (official)

  A stated review budget is what makes a score comparable across tools, and it is the part a team can adopt without the benchmark: report how long a reviewer had, not just how often the model was right.

• [2026-08-09] Android accessibility trees are unsanitised text, and mobile agents read them as instructions. Injecting prompts into UI labels hijacked goals, drifted context and triggered unauthorised device actions: 0.822 attack success against MobileRun on Gemma4:31B, 0.150 against Mobile-Use on Qwen3.6:35B. Same attack, same phone — the framework’s input handling decided the outcome. (official)

  Any app that renders a label controls text the agent reads as instruction, so the untrusted boundary on a phone sits far lower down than the pages an agent visits.

Watch list

• Qwen3.8-Max weights, mid-week in the week Alibaba named. Neither the Max checkpoint nor the Qwen3.8-27B companion is on Hugging Face or ModelScope, and no licence has been stated. Qwen 3.5 and 3.6 shipped Apache-2.0, which is a pattern rather than a promise. The API model has been live since August 3, so a LICENSE file is the whole remaining question.

  A drop matching the last two licences would barely be news; what is worth watching for is one that doesn’t, since that is what would change how teams plan around the Qwen line.

• DeepSeek’s price rise, five days on, still a sentence with no number. No rate, no effective date, no word on whether V4-Flash, V4-Pro and cache-hit pricing move together. Current published rates are $0.14/$0.28 and $0.435/$0.87 per million tokens; until a new card replaces them there is nothing to migrate against, only a fallback worth pricing.

  Whether cache-hit pricing moves with the headline rate is what decides this for batch workloads, and it is the detail a one-sentence announcement is least likely to include.

• Agent Plugins 1.0 still has no install path. A week after the format landed with Google, Amazon, Cursor, Microsoft, OpenAI and Vercel behind it, no client has shipped a release that reads a plugin.json. One package landing unmodified in two clients, skills and MCP servers both live, is the test; a client changelog entry is where it would show.

  Skills and MCP servers both having to work is what makes this harder than an unpack step, and it is why a shipped client release counts as evidence when another maintainer announcement does not.

• CoreBreak’s Strands gap, day five. AgentCore, ADK and the Vercel harnesses are patched; the Strands Python harness SDK still is not, and AWS still leaves the mitigation to callers. A working week of silence now reads as a decision about scope. A release note or a documented guard in the repo is what would close it.

  Nothing so far distinguishes “not yet” from “not planned”, and those lead to opposite decisions for a team currently carrying its own guard — which is why documentation would help even if no code shipped with it.