Open-source FPGA accelerator runs modern LLMs with full toolchain
The openTPU project releases a complete AI accelerator design in one repository, running models like Qwen3 and LFM2.5 on a Kintex-7 FPGA card with bit-exact simulation.
Daily coverage of AI, developer tools and infrastructure. Each story explains what happened and why it matters, with a link to the original source.
The openTPU project releases a complete AI accelerator design in one repository, running models like Qwen3 and LFM2.5 on a Kintex-7 FPGA card with bit-exact simulation.
Rapid7 reports new Linux implants BPFDoor and AVERAT disguising traffic as SpamSniper and ShareTech to evade detection in Asian telecom networks.
Researchers introduce Dust, a zeroth-order optimization algorithm that perturbs activations instead of weights to train transformers without backpropagation, showing competitive performance at scale.
Google releases EmbeddingGemma 2, a lightweight open model that unifies text, code, audio, and video embeddings for efficient on-device retrieval.
ProjectDiscovery demonstrates how easily attackers can embed hidden triggers in open-weight models to exfiltrate secrets, bypassing standard safety benchmarks.
Kubernetes v1.34 reaches General Availability for node swap, allowing clusters to use fast NVMe SSDs to page out idle memory and significantly increase pod density for AI workloads.
Citrix released urgent updates for CVE-2026-88779, a memory overflow flaw in NetScaler ADC and Gateway that attackers exploit to crash SAML deployments.
Reflection AI has unveiled Beam, an open-weight model claiming to match Chinese rivals in reasoning while using significantly less inference compute.
The open-source tool Rashomon independently records every command and subagent action in Claude Code, flagging discrepancies when the agent claims success despite underlying errors.
A new Google Research report argues that AI agents must understand social norms and context to ensure privacy and security, moving beyond traditional permission models.
Pi 1.0 integrates MCP servers but keeps tool definitions out of the prompt, using a JavaScript sandbox to discover and execute tools only when needed.
OpenAI is rolling out invisible text watermarks for ChatGPT and Codex in the European Union to comply with the EU AI Act, using a method called textGrain.
Attackers are exploiting CVE-2026-61500 in Rejetto HTTP File Server to forge admin sessions and execute code. The bug stems from a predictable pseudo-random number generator.
Apple plans stricter controls for macOS Full Disk Access to mitigate privacy risks from autonomous AI agents, following recent security vulnerabilities in Meta and OpenAI apps.
Cloudflare has launched Clef, a 27B parameter model optimized for document extraction and routing. It processes text and images in a single pass via Ollama.