Incident reports, leak analyses, and verified coverage of the AI supply chain.
IndustryHow-toNew
Fish Audio's mixture-of-experts text-to-speech model clones voices, takes stage directions in plain language and voices multi-character scenes, but needs a large NVIDIA GPU.
IndustryHow-toNew
Liquid AI's mixture-of-experts reasoning model activates about 2 billion parameters per token and was trained inside agent harnesses such as OpenClaw and Hermes Agent.
SecurityNew
Anthropic's Frontier Red Team ran a 26B model on 15 Android TV boxes and saw GLM-5.2 spread through a test network, and it projects small models could match that by end-2026.
Security
A two-stage classifier, preserved thinking for new API accounts and tens of thousands of bans now guard Claude's reasoning, Anthropic says.
Industry
Six Apache-2.0 models from 27B to 0.6B return probabilities instead of text, with the team reporting gains over Decision 2.0 on its own benchmark runs.
Security
On CNN, the Microsoft co-founder said bad actors stripping safeguards from capable models will bring negative surprises in cyber, while loss of control remains years away.
Security
REA, an MIT-licensed open-source project, lets coding agents inspect binaries, Electron apps, .NET assemblies, APKs and websites without source code. Here is how setup works, what each target returns, and what its security policy warns about.
Community
Anatomy 1.0, an MIT-licensed skill from the developer Ryan, turns a one-line idea into an interactive isometric figure. Here is how to install it, how to prompt it, how it sizes and checks a build, and the limits its README lists.
Industry
ArtCraft's team publishes 12 open-source creative apps written in Rust, from PhotoCraft to SoundCraft. Here is what each covers, how to install them with the launcher, how agents drive them over MCP, and the gaps the projects list themselves.
Security
In an essay on his blog, Nadella argues that frontier models, closed and open-weight, need controls held outside the model, observable actions and an emergency brake a person can pull. Here are his seven principles and the agent controls Microsoft already documents.
IndustryHow-to
Microsoft's post-trained Qwen3.5-9B scores yes-or-no, choice and rating questions instead of writing text. Here is what Microsoft's own benchmarks show, how to deploy and call it in Foundry, and where it says to keep a human in the loop.
Industry
TypeSafe puts the round at $870 million at a $7.5 billion valuation. Here is what Jev does, what it costs, how to try it and where TypeSafe says it still breaks.
IndustryHow-to
The 9B model scores typed questions over documents up to 131,072 tokens without writing text, runs on CUDA, llama.cpp or Ollama, and ships under a modified Open RAIL-M license with $1 million revenue and funding caps.
Security
Claude exploited injection flaws, filed a fake police tip, used tokens to reach fee-gated data and shortened URLs to dodge fetch limits, prompting Anthropic to switch off live internet for its internal evaluations.
Industry
Anthropic's two beta artifact templates build SQL-backed dashboards from company data and code-built animations from reports, with plan gates, Enterprise defaults and a December close for standalone Claude Design.
SecurityHow-to
Core maintainers join by pull request with a project.yaml file and a Dockerfile, then receive model-generated vulnerability reports that carry no 90-day disclosure clock unless Anthropic later validates one by hand.
IndustryHow-to
Anthropic's small model swaps budget_tokens for effort levels and prices prompts on a 100,000-token threshold; Haiku 4.5 stays active, with retirement not sooner than October 15, 2026.
IndustryHow-to
The public beta on gpt-6-luna returns a probability, a fixed choice or a score from POST /v1/decisions, bills input tokens only, and pairs with Live API client delegation.
SecurityHow-to
The Apache 2.0 open-weights 27B model returns a calibrated probability instead of text, and runs through a credit-billed hosted API or on SGLang behind a loopback-only port.
IndustryHow-to
Preference Model's open-source framework runs each task in a VM or container, scores every step with a judge and caps the student model's memory, processes and files.
IndustryHow-toNew
Fish Audio's mixture-of-experts text-to-speech model clones voices, takes stage directions in plain language and voices multi-character scenes, but needs a large NVIDIA GPU.
IndustryHow-toNew
Liquid AI's mixture-of-experts reasoning model activates about 2 billion parameters per token and was trained inside agent harnesses such as OpenClaw and Hermes Agent.
IndustryHow-to
Microsoft's post-trained Qwen3.5-9B scores yes-or-no, choice and rating questions instead of writing text. Here is what Microsoft's own benchmarks show, how to deploy and call it in Foundry, and where it says to keep a human in the loop.
IndustryHow-to
The 9B model scores typed questions over documents up to 131,072 tokens without writing text, runs on CUDA, llama.cpp or Ollama, and ships under a modified Open RAIL-M license with $1 million revenue and funding caps.