Open Source

Open-source models can hide backdoors that arm on a chosen date

The Guardrails

Open-source models can hide backdoors that arm on a chosen date

Researchers have shown you can download a coding model, run it daily, and never see a thing — until a date you've never heard of rolls around and the model silently runs a command nobody asked for. Morgin's team proved it by training exactly that behavior into a small open-weights model and firing it inside OpenCode. Morgin researchers trained a time-release backdoor into a 2B coding model that stays inert on every normal day and detonates on one chosen date.

Hugging Face explores a sale at a $13B valuation

The Arena

Hugging Face explores a sale at a $13B valuation

Two Sunday stories about who gets to own AI's plumbing: the open-source model hub everyone relies on is quietly shopping itself, and Sam Altman is warning that fear could hand AI to a handful of powerful players. Hugging Face is exploring a sale that could value the open-source AI hub at $13 billion or more. Business Insider reports the company has brought in a bank to sound out potential buyers; talks are early and no bidder has been named.

A local open-weights model built a working license bypass in 30 minutes

The Guardrails

A local open-weights model built a working license bypass in 30 minutes

XDA's lead technical editor handed Qwen 3.8 27B — an open-weights model small enough to run on consumer-class hardware — a task he assumed needed a frontier model: reverse-engineering a commercial app's license check. It refused the initial jailbreak, then audited the authentication scheme, recovered a deliberately obscured cryptographic key, caught and corrected its own mistakes, and produced a working bypass — all in about half an hour, entirely offline. The test ran on a single Lenovo…

Deep Dive — Nvidia's $6 billion plan to beat China at open weights

The Arena

Deep Dive — Nvidia's $6 billion plan to beat China at open weights

For a company that historically made its money selling the shovels, Nvidia just bought a pickaxe. The chip giant is taking the extraordinary step of licensing the model-building software of the Paris startup Poolside and hiring 109 of the people who used it — a deal people familiar with the matter say Nvidia intends to turn into one of the world's most powerful open-weight AI models, built specifically to answer Chinese heavyweights like DeepSeek and Kimi K3.

Quick Hits — August 22, 2026

Quick Hits

Quick Hits — August 22, 2026

Three notes from the AI ecosystem: a space robot brain going open-source, a clever trick for cheaper long-context inference, and a vibe-coded app that keeps finding fans. SpaceClaw ("Space Lobster"): the first open-source space embodied-intelligence project. Three companies — group-intelligence world-model firm UltraBrain, commercial spaceflight outfit Future Aerospace, and flexible-arm maker Xiyun Robotics — unveiled SpaceClaw at this week's World Robot Conference, aiming to land a real…

Open Source Radar — August 22: swarms, workspaces & observability

The Stack

Open Source Radar — August 22: swarms, workspaces & observability

Today's board is the agent operating layer: a meta-harness for orchestrating teams of agents, Apache's new local-first agent workspace, and an analytics platform that wants to be every AI team's eyes on what agents are doing. Three projects, all fresh on daily trending, none of them repeats. ruflo (TypeScript, ~68k stars) — An agent meta-harness from the developer behind claude-flow, built to deploy intelligent multi-agent swarms and coordinate autonomous workflows from one place.

Alibaba open-sources Qwen-UI-Agent trained on 100 real phones

The Frontier

Alibaba open-sources Qwen-UI-Agent trained on 100 real phones

Two stories caught our eye this morning: Alibaba's bet on putting AI inside real devices, and a reminder that AI policy is now a campaign war chest. Alibaba open-sourced Qwen-UI-Agent, a GUI agent model trained and evaluated inside a fleet of more than 100 physical phones rather than clean simulators. The model is built to drive devices the way a person would — tapping, typing, running command-line actions, and calling APIs to finish multi-step tasks across mobile, desktop, and web.

SenseTime open-sources SenseNova U1.5-Lite multimodal model

The Frontier

SenseTime open-sources SenseNova U1.5-Lite multimodal model

Two Chinese outlets and the project's own repository confirm the same move: SenseTime has pushed its lightweight unified multimodal model, SenseNova U1.5-Lite-Preview, into the open. The model lands as the multimodal race shifts from raw parameter scale toward efficiency, fine-grained control, and whether a smaller model can actually do useful work. SenseTime open-sources SenseNova U1.5-Lite, an 8B unified multimodal model with native 4K image generation.

Mojo compiler goes open source under Apache 2.0

The Stack

Mojo compiler goes open source under Apache 2.0

Modular has thrown open the source for Mojo, the systems programming language it built for the AI-accelerator era, releasing the full compiler and toolchain under the Apache 2.0 license (with LLVM exceptions) on GitHub. For a language that spent four years in public development behind a closed compiler, this is the unlock that turns Mojo from a vendor experiment into something the wider community can actually own. Modular has open-sourced the entire Mojo compiler and toolchain under Apache…