Thinking Machines ships Inkling 2 Mini: a 26B looped MoE with Apache weights, built for 32GB machines
Inkling 2 Mini runs 3.4B active of 26.4B params, loops its middle layers three times, takes 1M tokens and fits a ~20GB checkpoint. Apache 2.0.
298 sightings logged, newest first.
Inkling 2 Mini runs 3.4B active of 26.4B params, loops its middle layers three times, takes 1M tokens and fits a ~20GB checkpoint. Apache 2.0.
Amazon says a China-linked actor's Cairn agents found a file-write zero-day in a medical imaging AI app in 11 minutes. MONAI Label has no fixed release yet.
Kubernetes from Zero · Foundations 9 of 12 · Lesson 9
Inject settings into Pods as variables and files, see which ones update live, learn why a Secret is only base64-encoded, and lock a ConfigMap as immutable.
F6 reports the first in-the-wild attack on clustered Cisco Expressway via CVE-2026-76466 (CVSS 9.8). Cisco still lists no known abuse. Patch to X15.5.20.
The two open-weight image models to know: MiniMax H3-Image for 16 MP and 16 references, HiDream-O1 for an MIT licence at 8B. Ranks, hardware and licences.
A new preview in the Gemini Interactions API keeps big tool libraries out of the prompt until needed. No extra charge, and it works with remote MCP servers.
Microsoft's new coding models are in Copilot now, with cost claims from Microsoft's own tests. Decision-1.1, its router model, also went GA.
Mandiant says a China-linked group used vCenter auth bypass CVE-2026-123890 from early September. Broadcom's fix has no workaround; 7.0 has no public patch.
Kubernetes from Zero · Foundations 8 of 12 · Lesson 8
Give Pods a stable name and IP with a Service. Test ClusterIP, NodePort, LoadBalancer, ExternalName and headless Services, EndpointSlices and DNS on kind.
Two rogue agent-framework releases ran a stealer on every Python start, taking cloud keys and CI tokens. Pin 1.26.0, check for .pth files, rotate secrets.
PixVerse V6 is a cheap 1 to 15 second clip model with up to 10 references. HeyGen Video runs to 30 seconds and leads one AA board. Checked 19 Nov.
CVE-2026-114938 corrupts the heap in OpenJDK's AVX-512 Base64 decoder. Calif shows pre-auth RCE in Jira and Jenkins. Patches out; probing has begun.
Prime Intellect's Model Factory beta runs RL, SFT, pre-training and one-click endpoints for open-weight models, per GPU-hour or on dedicated clusters.
Hugging Face Inference Providers now send LLM calls via OpenRouter, with retries and :fastest or :cheapest routing. Your own provider keys no longer work.
nginx 1.30.7 and 1.31.9 fix 10 CVEs, two of them HTTP/2 memory bugs Wiz rates 9.2. If you can't patch yet, turn HTTP/2 off. F5 says no exploitation seen.
Google's Gemini 4 Neon sits between Hydrogen and Carbon at $3/$15 per million tokens, with a 1M window. Live in the API, Antigravity and GitHub Copilot.
Kubernetes from Zero · Foundations 7 of 12 · Lesson 7
Run Pods through a Deployment: self-healing, scaling, a rolling update with maxSurge and maxUnavailable, a stuck rollout and a rollback, all tested on kind.
OpenAI lists GPT-6.2 Astra at $8 input, $0.40 cached input, and $40 output per million tokens. GPT-6.2 Sol keeps GPT-6.1 Sol's $2, $0.10, and $10 rates.
Kubernetes from Zero · Foundations 6 of 12 · Lesson 6
Organise a cluster with namespaces, a context namespace, the app.kubernetes.io label set and annotations, and learn which objects live outside any namespace.
Use indeterminate progress when work is underway but its extent cannot be measured.
A lazy merge combines existing order; it does not repair a source whose records arrive out of order.
Both can appear in the top layer, but their interaction and dismissal rules differ.
An overlay can be concise while still discarding a nested collection from the earlier configuration.
Kubernetes from Zero · Foundations 5 of 12 · Lesson 5
Imperative vs declarative kubectl, YAML from dry-run, apply, diff, explain, edit and patch, every -o output format, and label selectors, all tested on kind.
Claude Code 2.1.323 adds Code Mode, an opt-in where Claude writes JavaScript to run several tool calls in one step. Off by default, terminal only.
Choose a useful starting point and make the return journey work for keyboard users.
Keeping dict.keys or dict.items does not freeze the dictionary at that moment.
Prices rechecked 14 Nov. Hume's old TTS docs now redirect, Chatterbox's README calls the repo an archive, and the main API prices still match.
Kubernetes from Zero · Foundations 4 of 12 · Lesson 4
Write a Pod manifest, apply it to a kind cluster, then inspect it with get, describe, logs, exec and port-forward, read its phase and delete it.
Pause optional work when a document is hidden, and resume carefully when it becomes visible.
Line history starts from a particular revision and follows the selected range, so choose that starting point deliberately.
Project Zero ends the extra 30-day patch soak. Disclosure is 90 days from 13 Nov 2026, 60 days from 2 Feb 2027, and 30 days from 3 Jan 2028.
US Gemini app users can connect Google Wallet. Bank and card links through Plaid are optional and read-only. Disconnect in Connected Apps.
Bitrise is back online in a limited safety mode after its SEV-1 incident. Sign in and rotate every stored secret, token and certificate.
IntersectionObserver reports geometry, not whether a person saw or read an element.
Pickaxe occurrence counts and matching changed lines find different kinds of edits.
Copilot deprecates Claude Fable 5, Haiku 4.5, Gemini 3.8 Flash, Gemini 4 Argon, Grok 4.6 and Kimi K3 on 14 December. Here are the suggested replacements.
Kitesurf, Cloudflare's stateless browser for AI agents, is now on GitHub under Apache 2.0. Run it locally over CDP or deploy it to your own Workers account.
Microsoft's MAI-Image 2.6 and 2.6 Flash, plus the editing-only 2.7 preview: resolution, references, prices, AA ranks, licence terms and who should use them.
Kubernetes from Zero · Foundations 3 of 12 · Lesson 3
Install kubectl and kind, compare minikube, k3d and Docker Desktop, run a three-node cluster, and learn how kubeconfig contexts pick a cluster.
ggml-org tags llama.cpp and ggml v1.0.0 and promises major-version bumps for any breaking change to the public C API, which helps every app built on them.
A size callback can create feedback. Make writes converge before reaching for animation frames.
PostgreSQL's 12 November update closes 52 CVEs across 14 to 18, ships the last PostgreSQL 14 release, and asks some users to reindex or take a new full backup.
Git can flag whitespace problems, but the comparison you select determines which changes it sees.
Vercel Labs' UI SDK beta ships 50+ behaviors like popups, focus and typeahead that agents compose into components for React, Vue, Svelte, Solid and the web.
Cognition gives every Teams user two free SWE-2 sessions, makes free SWE-2 permanent on Pro, Max and Teams, and adds SSO, cost tiers and a standard BAA.
DeepSeek's 13 November API release adds server-run web search, file parsing, URL reading, TTS, ASR and a safety check to deepseek-flash, with per-use prices.
Microsoft says servers failing automated health checks cut Exchange Online access and broke Copilot. It found no sign of an attack and promises a report.
Google's Gemini 4 Hydrogen costs $1.50/$7.50 per million tokens. In Antigravity, Orchestrator pairs it with Argon on Pro and with Carbon on Ultra.
A race in mod_http2 lets an unauthenticated client corrupt httpd memory and run code. Upgrade to 2.4.71 or turn off h2 and h2c until you can.
A custom error stays in place until your code clears it, even after the value changes.
A stable list needs a final ordering key that distinguishes otherwise equal rows.
LiquidCodec turns 24 kHz speech into 12 tokens per 80 ms frame and back, loads in Transformers, and uses Liquid's revenue-capped LFM Open License.
Arm's Lorenzo Stoakes posted patches removing himself as reviewer for MM core, misc, reclaim and Rust, saying AI-generated series left him no bandwidth.
Server, Postgres session store, sandboxes and credentials can now run in your cloud or on-prem. Model calls can go to Anthropic, Bedrock, Google or Foundry.
Kubernetes from Zero · Foundations 2 of 12 · Lesson 2
Name every control-plane and node component, then follow one kubectl apply step by step until a container runs.
VS Code 1.146 deprecates the Local agent harness, which goes in February 2027. It also adds an experimental session Inbox and agent-run model comparisons.
SubmitEvent.submitter identifies the initiating control when there is one.
Daybreak Red Team sends a team of agents at your web apps and is in early access. Defense Factory turns findings into patches that engineers approve.
Combining two result sets needs a decision about whether repeated values represent repeated events.
Eight changes for people building with AI: an active CI stealer, a GPUStack patch, a Val Town shutdown and three new tools.
CoreWeave's managed PostgreSQL runs on CPU nodes in your CKS cluster beside your GPUs, with HA, point-in-time recovery and pgvector. Built on CloudNativePG.
Microsoft FORGE Lab's MDASH scan of Linux took two weeks. Turning it into 59 kernel reports took months, and the oldest bug was at least 21 years old.
From 90 to about 1,300 tokens per second per GPU on 384 GH200s: how the EUROPA team fit a 952B MoE onto shared EuroHPC hardware with PyTorch and TorchTitan.
A lab demo, not a Claude flaw: with a hacked machine and a stolen API key, Darktrace steered Claude Code through memory stores and pulled out a 1 GB file.
India joins Australia and Japan for in-country API inference on gpt-6-luna. It needs Zero Data Retention, Standard pay-as-you-go and a 10% residency uplift.
OpenShell, NVIDIA's runtime that enforces what AI agents may touch, is now a CNCF Sandbox project. It moves to a neutral ai-openshell GitHub org.
DuckDB's team explains why storing data by column makes it fast. We tried the ideas on a small LLM eval log, with short queries you can run in a minute.
In four simulated tasks, one agent serving everyone beat teams of per-user agents from Anthropic, OpenAI and Alibaba. Three environments are out as MAMUBench.
A critical use-after-free in MongoDB Server's replication code is reachable without credentials. Fixes are out for 4.4 through 9.0. Patch every member.
Request a normal submission when validation and submit handlers matter.
LTX-2.5 is open weights with per-second API prices up to 4K. Vidu Q4 Preview takes 15 reference images and ranks higher on image-to-video. Checked 12 Nov.
Rows with the same ordering key can share a cumulative result unless the frame says otherwise.
Since 10 November, stolen GitHub logins have opened probe PRs in 60,000+ repos, then pushed a stealer that grabs CI secrets and deletes its own run logs.
ACE Studio's ACE Music v1 leads AA-Music-Vocal at 1160 Elo to Suno v6's 1144, but the 95% intervals overlap. Browser app only; a Flash model is due open.
Google's 10 November Antigravity post rolls Gemini 4 Carbon into the App, CLI and IDE extensions for AI Ultra, and previews /swarm on Pro and Ultra in the App and CLI.
PAN 2 pairs a language model that plans the next world state with a 29B video model that renders it, up to 20 seconds. Weights and code are Apache 2.0.
Kubernetes from Zero · Foundations 1 of 12 · Lesson 1
Why containers led to Kubernetes, what desired state means, and your first cluster answering on your laptop with kind.
United says personal AI agents may no longer book, buy or manage trips on united.com, in its app or through travel agencies without its permission.
Autocomplete tokens distinguish address purposes and group repeated fields.
AWS Emulator is a free, CC0 local stand-in for AWS APIs, written in Rust. AWS says autonomous agents wrote and maintain every line, with no human review.
Explicit NULL ordering makes a ranked list easier to understand and maintain.
Velo Research open-sourced Kinetic, a Rust drop-in for llama.cpp's tools. Its big speedups are on file-rewrite tasks; plain generation is about level.
OpenAI drops Advanced and Standard voice from ChatGPT on 11 December 2026. GPT-Live takes over; live video and screen sharing stay on Advanced for now.
Relate is an Apache 2.0, in-process graph database in Rust with Cypher, vector search, a CLI and an MCP server. Neo4j calls it an experimental Labs project.
AIR Security says Anthropic's docx, pptx and xlsx skills load a LibreOffice shim from a fixed temp path without checks. Anthropic calls it out of scope.
Air ships in all 2026.3 JetBrains IDEs. Agents get inspections, builds and the debugger through MCP, and a paid JetBrains AI plan is not required.
Praxis-1 weights, inference and fine-tuning code are public under Apache 2.0. The quick start needs one 80 GB GPU; Runway's benchmark numbers are its own.
Gemini Decision 1 answers up to 1,000 yes/no, choice or rubric questions per call over text, images, audio, video and PDFs, at $0.10 per million input tokens.
Autocomplete tokens distinguish an existing password from one a person is creating.
A dashboard count needs to say whether it measures records or populated fields.
A hidden identifier can carry context through a form, but the server must still authorize its use.
A uniqueness rule on an optional field does not also require that field to be present.
Val Town closed signups on 10 November, paused billing and goes read-only on 10 December. It has migration guides to Cloudflare, Vercel, Netlify and Render.
Businesses can now claim their ChatGPT profile, add product feeds and see mention counts. US ecommerce firms can also launch a Business Agent, in beta.
Google DeepMind's new robot model reports 30% to 100% higher task success and 96% on a Mercedes-Benz line. Only partners get it; the safety benchmark is open.
Siemens patched a 9.2 CVSS flaw where Mendix Studio Pro's MCP server was reachable from the network. A fast-agent advisory shows the same default can mean RCE.
Anthropic adds shared whiteboards and spreadsheets to Claude on paid plans, renames Dashboards to Claude Data at GA, and says Claude Meetings is coming soon.
Vates says attackers are exploiting a critical Xen Orchestra login bypass that can lead to root on XCP-ng hosts. Update to 6.9.2 or 6.10.1 now.
The FMF's v1.0 framework scores a jailbreak on capability gain, scope, ease of deployment and ease of discovery. Public comments are open until 21 December.
Gemini for healthcare adds clinical and payer skills plus EHR, claims and coding connectors to the Gemini agent. It is in preview with no published price.
Mittwald says attackers who used its MCP server to reach 56 projects are back. Its website was hit and it shut its customer panels and SSH on 12 Nov.
MiniMax H3-Image makes and edits images up to 20 megapixels with up to 16 references. The API is live and the weights are out under a restrictive licence.
AI agents wrote about 6.5 million lines of Lean proving that libsignal's compiled AArch64 code keeps Signal's security promises. Humans approved the specs.
Nuance Labs launched Emma on Nuplex 1.0, a full-duplex audiovisual model that listens as it speaks, with a beta API that bans cloning and adds watermarks.
RFC 10008 standardizes QUERY, a safe and idempotent HTTP method with a request body. It is the clean way to send a large or private query.
An SGLang remote code execution flaw let an attacker into NEAR AI's inference servers, about 100 accounts and 18 agents. Confidential mode was in scope too.
Voxtral Mini TTS 3 is $13 per million characters with an agent SDK; MAI-Voice-2.1 is $22, Flash $15, with gated cloning. How the two compare, checked 10 Nov.
ZONOS3-Mini is an 800M-parameter MoE text-to-speech model with about 100M active, Apache-2.0 weights, voice cloning and runtimes for browsers, Macs and SGLang.
Guide people toward expected files with accept, then validate uploads on the server.
GPUStack before 2.2.0 hands admin to anything that looks like localhost, including traffic through a same-host proxy. CISA says it is exploited. Patch now.
Choose the comparison mode and cache policy before treating a matching result as proof about file contents.
OpenAI's Control Room makes every research inference engine ask a central service before answering, so a stopped job gets refusals. It fails closed if down.
After the OpenAI agent incident, Australia plans binding rules for frontier AI firms: a new regulator, rapid incident notice and independent model checks.
serc is an experimental Rust compiler for CPython bytecode that claims byte-identical output 3-4x faster than compileall. Every code change was by GPT.
Soft and hard wrapping can submit different text even when the controls look alike.
Fireworks' 9 Nov update widens its October incident: billing data, and for some accounts secrets, datasets and incomplete model weights. Inference unaffected.
A fixed three-part result makes a missing delimiter explicit without splitting the rest of the value.
Fable 5.2 is out on every major cloud with a 1M context and fast mode. Anthropic says it beats Opus 5.5 on most tests. Mythos 5.2 stays gated.
The 27th tutorial: set destination allowlists, turn on Action Recording, and stream scrubbed Bot actions to your collector.
Exact, an MIT-licensed Expo project, runs one Contract UI natively on web, macOS, iOS and Linux, with a ten-command driver for agents. It is pre-alpha.
GR00T 2 learns robot skills from 400,000+ hours of human video. NVIDIA reports 40% zero-shot on unseen tasks, 95% after five demos. Code not out yet.
Managed Postgres is GA on AWS, with backup storage and egress now billed. Xata's copy-on-write branching and scale-to-zero are coming to it; no date given.
Nitro Sandboxes are GA in all commercial AWS Regions at no extra charge. AWS pitches them for running LLM-generated code and agents on shared EC2 instances.
Muse Ember is Meta's smallest model: $0.10 in and $0.50 out per million tokens, 1M context, API only. It ties Claude Haiku 5.5 on Artificial Analysis's index.
Bedrock AgentCore's Web Answer plans its own searches and returns a synthesized answer with citations. us-east-1 only, $10 per 1,000 calls, no model choice.
Ferrite is a Rust LLM engine for DGX Spark with Rust GPU kernels. Since 0.1.0 it added Intel Arc Pro B70 and 16 GB RTX targets. Experimental, Apache-2.0.
HOPE scores experts in pairs, not one by one. Its CLI writes a standard Hugging Face checkpoint. AWS reports 42.3 vs 41.8 at 50% pruning on Qwen3.5-35B-A3B.
A multi-year Adobe and Disney alliance: custom gen-AI models proprietary to Disney, starting with Imagineering's sketch-to-image tool. No terms disclosed.
strutil-core 1.0.1 stole credentials on install and planted MCP servers, hooks and AGENTS.md files so coding agents would rerun it. Pulled within hours.
Two more eighth-gen TPU systems, a TPU agent suite for coding tools in preview, GKE dynamic slicing GA and an Apache-2.0 C++ serving runtime. No prices yet.
CHIMERA speaks, sings and beatboxes in the browser from a simulated airway, in Rust compiled to WebAssembly. Its own Whisper check: 37% word errors vs 2.7%.
Official Hermes plugin threat-models a local repo, reviews it with parallel subagents, can test findings and exports SARIF. Your model sees the code.
Microsoft's azd Foundry extension is GA at 1.0: hosted agents live in azure.yaml, test locally, deploy with azd up and gate CI on evals via GitHub Actions.
Tan's open harness evals held Opus 5.5 and GPT-6.1 Sol fixed across seven setups. Pass rates could not be told apart; cost per solved task ran $0.17 to $0.44.
Torvalds' 7.4-rc2 note says much of the work was AI-driven fixes and relays Alexei Starovoitov's claim that BPF fixes now arrive faster than new reports.
Textarea length limits use UTF-16 code units, with important validation and newline caveats.
ByteDance Seedream 5.0 Pro and today's 5.5 against Alibaba Qwen-Image 3.0 and open 2.1: resolution, refs, prices, AA ranks, licences and who fits which job.
Text line handling needs an explicit boundary rule, especially when the source is a protocol.
9to5Mac and MacRumors quote Gurman: Apple could make Siri more like an agent next year, and Ternus pushed Federighi to move faster. Apple is not quoted.
Use browser network indicators as hints, then check the service you actually depend on.
Botpress says a staff API credential read bot configs in 16 workspaces on 21 Oct, 07:44 to 08:23 UTC, and says to treat those stored credentials as exposed.
Register each successful acquisition so a later failure can unwind the earlier resources.
MIT-licensed GPU Taskman v0.3.3, released 7 November 2026, reports per-process VRAM and estimates LLM weights and KV cache.
EU and EFTA tenants can enable Claude in the Microsoft 365 admin center. Boundary processing is dated to the end of 2026.
AP cites two unnamed former officials who say CAISSI staff were asked to informally report to a White House office.
Seedance 2.5 supports up to 50 references. On Artificial Analysis's 8 November audio boards, it leads image-to-video and Wan 3.0 leads text-to-video.
Check HTTP status separately from network and request failures.
Copying the outer dictionary does not isolate the lists and dictionaries inside it.
HyperFlow H3 Flash is Video Rebirth's few-step MiniMax-H3 for joint video and audio. The H3 community license still sets the territory limits.
On 5 November ASD's ACSC published two papers on open-weight models: inventory them, check hashes and signatures, and approve risky actions outside the model.
A fulfilled fetch promise does not mean its response body has finished arriving.
Anomaly's Rune is a pre-release MIT UI kit an agent can render to PNG. Voyager's vork 0.2.0 is source-available, and its license says it is not open source.
Memoization changes object lifetime as well as how often a function runs.
CVE-2026-116635: before 0.2.1, a crafted project name let an influenced agent read JSON outside its folder, or overwrite it with writes on. Upgrade.
LiteLLM merged a Rust router core, deleted the Python router and moved proxy Redis traffic to Rust. It ships in v1.110.0-dev.1, not a stable release yet.
Browse pairs an MIT-licensed MCP server with a Chromium extension so agents use your own tabs and logins, and hand 2FA and payments back to you.
PSI published a Lean-checked proof of a 2008 quantum query conjecture, found by its system on an unnamed Anthropic model. Experts are still reading it.
Hugging Face's harness adapters are LoRA weights trained with RL inside unmodified pi, opencode and hermes, with safeagents vetting every tool call.
Nous fine-tuned Qwen3.8-27B on 52.8M tokens and says propaganda answers fell from 22% to 0.9%. Apache-2.0 weights, eval and data are public.
A nested iframe in ChatGPT's HTML preview let pages open unsandboxed popups, defeating Lockdown Mode. Zero-click in ChatGPT Classic. Fixed 18 Sept.
Google measured TODO(bug) comments derailing its internal coding agents: unprompted detours cost about 70% more steps and 77% more tokens.
Linkerd's maker released Kinference: Apache-2.0 Rust control plane for vLLM on Kubernetes, with KV-cache-aware routing and policies. It is alpha.
Apiiro found typed passwords and keys saved in .claude/settings.local.json in public repos. How the file works, and the cleanup steps it recommends.
Linum's P-JiT 21K is a 2.18B pixel-space image model trained on 8,415 H100-hours. Apache-2.0 weights and code; it cannot render text.
Scope asked 20 agent setups to sign up for 18 hosted databases and use them. Prisma Postgres led at 88.3%; seven products had no successes.
Three open-weight TTS picks: Kokoro's tiny Apache model, Resemble's MIT Chatterbox family with cloning, and Fish Audio Drama 3, the top open model on AA.
Rivet's open-source Durable Objects alternative now has a GA Rust SDK (rivetkit) with SQLite, WebSockets, cron, hibernation and a test harness.
Cancel discards the remaining stream data and requests source cleanup. Releasing a lock leaves the stream for another reader.
Human-guided DeepMind agents found a faster signature-forgery attack and cheaper batch factoring for 1024-bit RSA. RSA-2048 is not affected.
A fixed-length recent-history buffer is useful precisely because it does not preserve every entry.
HashiCorp's GA Terraform policy agent turns Sentinel and OPA rules into Terraform policy with tests and reports, using Copilot, Bedrock or LiteLLM.
OpenAI and Ligo built 100 wet-lab tasks from raw gels, plates and microscopy. GPT-6 Astra gets 66.7% at xhigh; OpenAI lists where it still fails.
Five fal-served video models roll into Microsoft Foundry as a first-party Azure service in public preview, priced per generated second.
Anthropic's engineering post shows Claude reorganizing memory between sessions. The Managed Agents Dreams API is a research preview.
Google Agent Substrate 1.0 is GA on GKE; Substrate, gVisor and OpenRL enter CNCF Sandbox. Vendor claims sub-500ms restore and ~10x density.
Reading a response consumes its body stream. Clone before the body is locked or disturbed when two consumers need it.
AWS sample: Claude Code tasks on Bedrock AgentCore microVMs, Guardrails on prompts, short-lived GitHub tokens; the agent opens PRs and never merges.
Splitting an iterator can retain a backlog when one consumer runs ahead of the other.
Public repos lose pull_request_target by default on 2 December, with brownouts on 9 and 16 November. Actions lockfiles that pin action commits are GA.
The 26th tutorial: route one desktop, or install your networking client on every hosted computer so Bots reach private services.
GOV.UK factsheet: Cyber Security and Resilience Bill may add agentic loss-of-control resilience measures for regulated entities. Not law yet. Soft Legal, UK.
Underdog's Woof TTS: ~31.7 MB / ~30.7M params, WebGPU in-browser demo at underdog.ai/escape. License CC BY-NC 4.0. No confirmed Hub GA.
Antigravity Remote Agents (v2.29.0): Google-managed cloud VMs for Pro/Ultra personal accounts; attach GitHub, Drive, uploads; CLI agy --remote-agent.
Inkling 2 Small: 594B MoE / 17B active with vision, on Tinker. AA Index about 31. Apache-2.0 weights now public on Hugging Face after safety work.
Paid Cursor plans with cloud agents: scaffold a web app, optional Postgres, Cursor SSO, security checks, then publish a shareable URL.
Dwarkesh Patel's Clocktower Bench: 386 games of Blood on the Clocktower; Astra leads at 1859 Elo; detection leads OpenAI's edge.
AItonomy's Science's Last Exam: 113 tasks across 9 disciplines; Opus 5 + Claude Code High 37.50; end-to-end still 0 for listed models.
METR's Oct 2026 Anthropic assessment: AECI 1.9× since Oct 2025, subjective ~2× pace, compute not the driver, majority of key evidence redacted.
Recraft V4.1 family against Krea 2 Large and Krea 3 editing: resolution, refs, API prices, AA ranks, ownership and who fits which workflow.
Meta is rolling out voice and expressive avatars in the Muse app in the US first, on mobile, web and Mac.
Perplexity Teammates are always-on work agents on Computer for Max subscribers on the web, each in an isolated env.
NVIDIA posts four Cosmos 3 Nano open-weight checkpoints for interactive driving and robot world simulation under OpenMDW1.1.
Keep decoder state across byte chunks so a split UTF-8 character survives intact.
Claude Personal Finance is a US beta on Pro, Max and Ultra: read-only bank links via Plaid, up to 24 months of transactions.
A slice of a stream consumes its source; a second reader starts after that work.
GPT-6 Cyber is Critical under OpenAI's Preparedness Framework. Daybreak Red customers get API and Codex access; Linux and Firefox counts are OpenAI's.
The Justice Department forms DOJ-SITF under Woodward and Weiss to coordinate SI enforcement with the White House task force chaired by Jay Clayton.
Dropbox files an 8-K for a ~$350M cash deal for Tigris, expected to close in Q4 2026, and launches S3-compatible DBX Agent Cloud.
Use inputmode to make text entry easier, then validate the actual value separately.
Google's video model is generally available with 3-15s clips, reference conditioning, draft-then-upscale, and a 30% price cut at 720p.
Sorted input and an equality check turn a useful boundary search into a correct membership lookup.
The 25th tutorial: widen your Cursor SSO app, let the Bot's Linux computer pass device-trust rules, and recreate or clean up computers in bulk.
JetBrains' Rust IDE 2026.3 includes Air agents in-IDE, Junie Lite, local models, and an optional debugger MCP with a rustrover-debugger skill.
Acronis TRU says a lone operator wired 12 AI agents and MCP hacking tools into a mesh that hit about 10 targets. Commercial models refused to attack.
AlloyDB Express gives developers and coding agents a 1 or 2 vCPU Postgres 18 database in seconds, scales compute to zero and has a 12-month free tier.
A researcher says any extension that controls claude.ai could make Claude Desktop run commands through Dispatch. The Register says Anthropic has not fixed it.
Viggle's Remake adds two LoRAs to MiniMax H3 to swap people, objects, weather or style in existing video. The weights carry H3's licence, which bars US use.
Eight changes from our past week for people building with AI: two security fixes, a new default, three deadlines and two new tools, each with a next step.
Hindsight 1.2 can hand small memory judgments to hs-decide-1-flash, an Apache 2.0 0.8B model. Vectorize says some operations get 2.5x cheaper; it is opt-in.
Google says its Project Suncatcher prototype ran Gemma 4 on a data center class TPU in orbit. Next: radiation and heat tests, then two satellites in 2027.
Kimi Hosted Agents runs the agent loop, sandboxes and session state for you on Moonshot's China platform, with 140+ skills, MCP and an enterprise input hook.
A real-robot benchmark from NUS, Hugging Face, BitRobot and partners: blinded A/B trials, every run on the Hub, and 10,000 free evaluation hours.
Baidu released a 280B-total, 14B-active pre-training checkpoint on Hugging Face. Its custom license bars use in the EU, UK and South Korea.
Render's agent sandboxes run as long as needed, suspend when idle and restore memory from snapshots. Oregon only, with key brokering in early access.
Hy4 adds image understanding and computer use. Tencent has now published the 770B weights under a custom license that limits big API resellers.
Check Point calls it prompt adware: a Codex jailbreak bundle that rewrites agent instructions so a greeting returns an ad for its author's LLM gateway.
A source-available Rust engine for Qwen GGUF models. Its author measures 113 tok/s decode vs 66.5 for tuned llama.cpp on an RTX 3060; prefill is level.
OpenAI says its pipeline validated and wrote candidate patches for 948 Linux local privilege escalation bugs. Some are fixed; keep kernels current.
An omitted button type can make a local UI action submit its form. Explicit types make the intended behavior clear.
A Personal Assistants category in Security Analytics, per-account agent views in early access, and per-agent Allow or Block in the renamed access policies.
A convenient insertion helper can still call an expensive constructor when the key already exists.
Two open-weight video models: MiniMax H3 adds stereo sound, 2K and up to 12 references; HunyuanVideo 1.5 runs on a 14 GB GPU. Prices, ranks, licences.
Ricursive put Ripple-65B Preview on Hugging Face under Apache 2.0. Cognition's SWE-2 Flash, built on an open base, withholds them. Scores are self-reported.
OpenAI is acquiring Pierre Computer Company, maker of Code Storage and the Trees and Diffs libraries already in the Codex app. The deal still needs approvals.
A new NCSC assessment puts AI-powered ransomware against UK critical infrastructure at 'highly likely' by 2028. On its yardstick, that means 80 to 90 percent.
Conductor, which runs teams of coding agents from Anthropic and OpenAI, is joining Meta. Export cloud workspaces before 20 November, when deletion begins.
Images 2.6 is in ChatGPT, Codex and the API as GPT Image 2.6 Flare and Sunburst. Token prices match GPT Image 2.5; the speed claims are OpenAI's.
Both can prevent editing, but disabled and readonly have different effects on form submission and validation.
Sarvam Prism returns a probability for every option instead of text. Sarvam says it scores 71.8 at 44 ms on a B200; local use needs a 120 GB FP8 GPU.
Why separated keys become separate groups, and when sorting would change the question you are asking.
The 24th tutorial is for IT: push the desktop app, open the two network paths it needs, and send each model request to your own policy check first.
OpenAI says Astra designed a device's enclosure, PCB and firmware with no direct human edits. Getting it manufacturable still took several rounds of review.
A Developer Preview in the 27.3 betas lets any app expose actions to Siri AI via App Intents. Shipping is months away; Siri AI is unavailable in the EU on iOS.
CVE-2026-111959: arrow-rs let a malformed RunEndEncoded array read past its values. Apache rates it moderate. Update to 60.1.0, 59.3.2, 58.4.1 or 57.3.2.
fal confirms a Bessemer-led Series F and a Sequoia-led Series E. Bloomberg reports $1.3B at $16B; CEO Burkay Gur says fal rate-limited users in summer.
Google's Antigravity team says an agent moved FVCOM's core solver into 13 safe-Rust modules in 59 minutes, and points to its migrate-to-rust skill.
Binance's initial report: a tampered Homebrew install of an AI coding tool, a malicious interview repo, then an internal AI tool used to reach engineers.
From 16 November, App Store Connect asks if an app has dating or a virtual companion with sexual or romantic content. If yes, the minimum rating is 18+.
P-H3-Image makes and edits images in 8 steps on MiniMax-H3. Pruna says the 4-bit set peaks at 21 GB of VRAM, but the license excludes the US, EU, UK and Korea.
A GitHub issue quotes Postman 12.33.4's own log installing Claude Code and Codex plugins, a startup hook and an MCP server. How to check your machine.
zg, a local hybrid search tool for coding agents from the zvec team, rebuilt its core pipeline in Rust. Its own M3 Pro test shows faster indexing, less memory.
Google says Gemini, Gemini in Classroom and Search will be off for K-12 student accounts until admins turn them on. Search comes with its AI features.
Playwright 1.64.1 still ships Chromium 156.0.8078.4, older than Chrome's 52-fix update, and runs it unsandboxed by default. What agent builders can do.
SolidCraft 0.30.1 fixes three advisories: an agent permission bypass, a lock-file symlink overwrite and an API key check skipped with '#'. Upgrade now.
Hydrogen-1.0 targets chemistry, materials and life sciences. The weights are MIT, but the GLM-5.3 base license has its own terms worth reading first.
CrowdStrike will buy Fiddler and Dreadnode, ships an AI Kill Switch and agent hooks for Claude Code, Codex and Cursor, and runs Falcon inside Cowork's VM.
Baidu says Apollo World generates live camera and lidar streams and reacts to the car's actions. It also proposed a '2+3' test for world models.
Paige and the renamed PD Reliability Platform are GA as of 3 November. Autonomous fixes are opt-in and marked coming soon; PD Labs previews are unsupported.
FlashProve attaches a zero-knowledge proof to LLM output so a user can check which committed model produced it. Prototype only: 1,024 tokens, ~70-95x GPU cost.
Stanford researchers retested 50 published LLM claims on 20 models: 44% expired, 34% weakened, median life 14.7 months. A project page, not a paper.
CVE-2026-112816 is rated CVSS 10. Canonical shipped fixed LXD and MicroCloud snaps on 29 October. Refresh, then check which servers listen on the network.
A non-binding opinion from Advocate General Szpunar in C-250/25 says chatbot answers can infringe press publishers' rights in some cases.
MIT CSAIL researchers say GPT-6 Astra lifted Qwen3.5-0.8B from 7.9% to 48.1% on MATH-500 by editing weights. A repro finds the gain is mostly loop removal.
KMS now offers ML-KEM keys, Aurora PostgreSQL 18 lets you pick post-quantum TLS groups, and a new CloudFront header shows which viewers use PQ key exchange.
Cartesia's Sonic 3.6 ranks 4th for TTS from $37 per million characters. Hume's Octave and EVI APIs shut on 13 Nov. Ranks, cloning rules and a migration plan.
Grafana 13.3 adds cross-dashboard variables, Notebooks and a UI refresh in public preview. Here are the toggles, the plugin impact and the one breaking change.
Azure WAF's HTTP DDoS ruleset for Front Door Premium is GA. It learns each profile's traffic and starts in Log mode. Its two geo rules are still in preview.
OpenAI says Pando, built in Rust with Codex in 10 weeks, cut median push latency by nearly 80%. It runs only inside OpenAI; you can't use it yet.
EKS Auto Mode now turns Gateway API resources into AWS load balancers: ALB for HTTP and gRPC routes, NLB for TCP and UDP. Ingress keeps working alongside.
WorkOS found unauthorized access to a sales tool on 30 October. CRM contacts and customer call and Slack excerpts were exposed; it says production was not.
Anthropic's message threads beta keeps a conversation server-side for 5 minutes, so agent loops send only new messages. Prices unchanged; not ZDR eligible.
The 23rd tutorial covers shared Bots: how to share one with your team, Collaborator or Manager, whose apps it uses, and who pays for each reply.
Microsoft patches the AKS app routing add-on's NGINX ingress only through November 2026. Its Gateway API successor gets a new IP, so plan a DNS cutover.
GitHub-verified MCP servers are in public preview for all Copilot plans: 26 services, no local install. Business and Enterprise admins must opt in.
EKS clusters created since 2018 have 10-year CAs nearing rotation. AWS rotates them, but nodes you manage and kubeconfigs must trust the new CA first.
Updated 9 November: Meituan released LongCat-2.5 weights on Hugging Face under MIT. Deployment docs cover NPUs only. API price was $0.30/$1.20 per million.
Trigger.dev blocked the exploit and emailed affected owners. It has not said how many orgs, how, or for how long. Rotate what you stored there.
Azure Functions ends support for the in-process .NET model on 10 November, the same day as .NET 8, .NET 9 and PowerShell 7.4. Microsoft points to .NET 10.
A preprint reports psychiatrists preferred the RL-trained agent in 76.7% of matched pairs and under-triage fell to 12.9%. No real patients were involved.
Rust's Niko Matsakis sketches evidence files that record what an agent checked and why, with CI gates and later re-audits. His crate is called habeas.
Pick Claude Agent, Codex, Antigravity or 40+ ACP agents inside Android Studio. Android skills, AppFunctions beta and A2UI for Compose round it out.
On Latent Space, Claude Platform leads describe a full Rust fork checked with shadow traffic. Ballpark: 40 to 50% less CPU, 10 to 20% faster first token.
Tencent says agent-created database instances grew over 80x in a year, with 1 million active, and added branching and faster serverless scaling.
A 9B multimodal model that stores 32 layers but executes 56, under Apache 2.0. Transformers only for now, with trust_remote_code. What the card says.
A Google-led paper builds a 4.1K-sample benchmark for implicit and changing requests. Every model tested gets worse as interactions run longer.
SkillSpector checks Claude Code, Codex and MCP skills for injection, exfiltration and supply-chain risk, then scores them 0 to 100. How to run it.
Vade runs Qwen3.8 Flash-Next on one to four DGX Sparks, writes its own GPU kernels in Rust and promises the same token IDs on every run.
From 7 November, text from grok-4.7 and other Grok API models carries an invisible watermark for the EU AI Act. Admins can opt out per model.
Steve Huffman says most scraping hits logged-out pages. Reddit will test login walls outside the US first; front page and shared links stay in question.
Rig's terminal agent now installs with cargo install rig. 0.46.0 adds opt-in MCP, ACP for editors, AGENTS.md, skills and traffic logs.
OpenAI says Plus, Pro and Business usage will be shared across ChatGPT and Codex and measured in tokens. Instant chat is excluded. No start date yet.
Chainalysis says names, emails, user IDs and role labels were accessed, with no sign of case data or tokens. What it said and what is still unknown.
Anthropic's table shows the watermark live on Mythos 5, Opus 4.7 and Sonnet 4.6 since 3 November. Opus 4.6 missed the date; a 16 November wave is reported.
Midjourney V8.2 against Ideogram 4.5 and 4.0: resolution, references, prices, ranks, open weights and who owns the output, on our 14-point checklist.
A file descriptor is just a number, so safe Rust code could hand any number to an API that trusts it. I/O safety closes that gap with OwnedFd and BorrowedFd. In October 2026 tokio deprecated four safe AsyncFd constructors for breaking it.
GitHub's hosted runners are thrown away after each job. Your own runner is not, and GitHub's own guidance says it should almost never serve a public repository, because any pull request can run code on it.
On Linux, load average includes tasks in uninterruptible sleep, waiting on disk or some locks. A load five times the core count can mean a machine stalled on memory or I/O, with CPUs that are not even full.
How a per-response nonce authorizes inline and external scripts, and how strict-dynamic lets a trusted script vouch for everything it loads.
The original and multilingual E5 embedding models were trained with a "query:" or "passage:" prefix on every input. Leave the prefixes off, or store documents with the query prefix, and retrieval quietly gets worse with no error anywhere.
How one response header tells a browser to erase its own cookies, cache, and storage for an origin when a session ends.
Some Rust crates fetch a prebuilt native library while they compile. The build works for months from a local cache, then a clean machine meets a refusing download host. Cargo's offline mode does not cover that download.
Why gzip and brotli belong to the representation itself, while chunked transfer is about how bytes cross the wire.
A Rust crate that wraps a C library declares it with the links key, and Cargo allows only one package per links value in a build. Two major versions of such a crate cannot share one dependency graph.
Why no-cache still allows caching and requires revalidation, while no-store forbids storing the response at all.
docker run -p 8080:80 listens on all host addresses, IPv4 and IPv6, and Docker's own docs say its rules bypass ufw. Writing 127.0.0.1 in front of the port is the difference between a local service and a public one.
How the Strict-Transport-Security header turns one HTTPS visit into a standing rule to use only HTTPS for as long as max-age lasts.
Files under /dev/shm and other tmpfs mounts live in RAM and swap. A df on the main disk never counts them, so a forgotten build folder there can starve a machine while every disk alarm stays quiet.