Perplexity's SPACE Red Team Exposes a Network Bypass Flaw Hitting 8 Sandboxes
Perplexity red-teamed its SPACE sandbox with nine frontier models given root access. None escaped the VM, but four bypassed network confinement.
20 stories
Perplexity red-teamed its SPACE sandbox with nine frontier models given root access. None escaped the VM, but four bypassed network confinement.
Eleven v4 tops the Provider Voice arena with a 1,319 Elo and hits 91.7% pronunciation accuracy, though it costs nearly 5x Google's Gemini Flash TTS.
Gimlet Labs is deploying 100 megawatts of Cerebras wafer-scale systems inside its multisilicon inference cloud, promising up to 3,000 tokens per second.
Shadcn refreshed his open-source chatbot starter with the latest Next.js, AI SDK, and shadcn/react, adding typed tool parts and a human-in-the-loop questionnaire flow.
Artificial Analysis launches a new benchmark measuring how well AI agents find and patch real security vulnerabilities, with surprising results on refusals.
A trending open-source framework chains reconnaissance, exploitation, and post-exploitation into one autonomous agent pipeline with human oversight gates.
A community-quantized, refusal-ablated text encoder for Qwen-Image-2.1 drops the refusal rate from 100% to 5% while running locally in ComfyUI.
H Company's new Holo4 family runs across GUIs, code, MCP, and APIs with a single model, competing with frontier systems at a fraction of the cost.
A new open-weight 320B Mixture-of-Experts model from IQuest Lab targets agentic coding in the terminal, activating 15B parameters per token with a 524K context window.
Strata is a purpose-built inference engine that runs Qwen3.8-Flash-Next, a 125B mixture-of-experts model, on a single gaming GPU with 64GB of system RAM.
Independent developer dzhng released jevgrep, a CLI that lets coding agents like Codex and Claude Code find code by asking what it does.
A new fix called GProj restores a broken conservation law in attention gradients, stopping late-training blowups that FlashAttention-3 silently causes in BF16.
A new study of 270 puzzles with multiple correct answers finds AI models cluster tightly together and diverge sharply from how humans pick solutions.
Two community LoRAs give MiniMax-H3 a full 360-degree camera orbit from a single photo, using opposite tricks with very different geometry results.
A new JEPA world model drives without a trained policy, hitting goals 20 meters out with a 100x inference speedup over prior latent planners.
InclusionAI open-sourced Ming-Image-0.1-Design, a 6B text-to-image model built for UIs, posters, and infographics with native transparent-background output.
Foreman pairs a Codex coding worker with TypeSafe's Jev classifier as an independent supervisor that scores nine yes/no questions in parallel.
Sakana AI and the University of Tokyo show that a vision language model can plan robot motions more reliably by testing and refining candidate trajectories in simulation before touching the real robot.
A 180B security-focused Mixture of Experts model is now available in eight local-friendly GGUF quants, with speeds up to 2.4 tokens per second on a single GPU.
OpenAI patched an image encoding bug that was quietly hurting vision quality on GPT-6 Sol and GPT-6 Luna across the API, Codex, and computer use.