Technology & AI
New LLM backdoor hides attacks in logical reasoning
An arXiv preprint shows language models up to 119 billion parameters can disguise backdoors as legitimate reasoning and evade safety checks.
AI Models
AI model releases and benchmarks — what each new model actually changes for the people who use it.
191 stories
Technology & AI
Together AI score centering stabilizes RL for LLMs up to 30B
Together AI's Marek and Ryabinin show score centering cancels drift in reinforcement learning of LLMs from 0.6B to 30B parameters.
Technology & AI
Voice agents drop 9.7% when instructions are merely implied
DSB-IFEval tests 1,038 cases across eight assistant roles and finds voice agents drop up to 9.7% when instructions are implied, not stated.
Technology & AI
When to trust LLM recommendations: four-tier framework
New 'epistemic warrant' framework grades individual LLM recommendations from unstable to broadly supported, capturing what confidence scores miss.
Technology & AI
New technique cuts AI agent wait time up to 45%
SMC pairs a small drafter model with a large actor model to speculatively execute tool calls, reducing wall time by up to 45% on AppWorld.
Technology & AI
Prompt injection attack on AI agents jumps to 28% success
SIR, a self-improving prompt injection attack, raised its success rate from 0% to 28% on Gemini 3.5 Flash while the agent's legitimate task still
Technology & AI
ToolSiphon drains 74% of data from LLM agent tools
New research shows ToolSiphon can extract most source records from LLM agent tools using only queries, affecting anyone building agent-based services.
Technology & AI
DiaSentinel AI agents screen diabetes risk on-premise
DiaSentinel, a new arXiv preprint from a Taiwan hospital team, describes a fully on-premise multi-agent system for type 2 diabetes risk screening from
Technology & AI
MMJailBench: prompt framing is top jailbreak risk across 16 AI models
MMJailBench, a new factorized benchmark, evaluated 16 multimodal LLMs and found prompt framing, not visual tricks, drives most jailbreak vulnerabilities.
Technology & AI
NVIDIA's $3.5B MediaTek deal extends AI from cloud to car
NVIDIA's $3.5 billion convertible bond investment in MediaTek locks in a partner spanning AI factories, consumer PCs and autonomous vehicles.
Technology & AI
LLM security agents act but lack guardrails, review finds
A systematic review of LLM-based security agents from 2023 to 2026 finds the field can act but cannot yet bound authority or audit behaviour.
Technology & AI
Osmantic ODS turns your PC into a private AI server
Osmantic ODS bundles Ollama, Open WebUI, n8n and ComfyUI into a one-command local AI server. Here's what the README claims and what's unverified.