Technology & AI
Bulkhead uses LLMs to detect and patch container escape bugs
Bulkhead, a new arXiv preprint, uses multi-agent LLMs to automatically find and fix path traversal vulnerabilities in containers running AI workloads.
AI Models
AI model releases and benchmarks — what each new model actually changes for the people who use it.
191 stories
Technology & AI
MindReader: LLM tool makes replacement passwords harder to guess
A new arXiv study finds LLM-suggested password replacements are more secure than human-created ones and equally memorable after one week.
Technology & AI
30 prompts fine-tune AI to near-optimal energy storage control
Reinforcement fine-tuning with 30 prompts cut an open-weight model's building emissions to 61.2 kg-CO2, near the 60.8 optimum, a preprint shows.
Technology & AI
LLM judge bias found in hidden states, steerable in both directions
Researchers find LLM judge bias lives in a low-dimensional subspace inside model hidden states, and steering along it can switch bias on and off.
Technology & AI
StoryTeller: training-free AI for film audio descriptions
A new arXiv preprint introduces StoryTeller, a training-free system that keeps film audio descriptions coherent for blind and low-vision viewers.
Technology & AI
Transformer learning dynamics reduced to few coordinates
New arXiv preprint shows Transformer training on inductive reasoning can be confined to a low-dimensional manifold for automatic circuit detection.
Technology & AI
LLM temperature setting steers ideological bias in RAG answers
An arXiv preprint finds sampling temperature on retrieval-augmented LLMs controls how strongly ideological framing from source documents bleeds into
Technology & AI
AI reasoning scaffold helps one model, hurts another
Reasoning scaffold lifts GPT-4.1-mini by 0.21 but degrades GPT-5-mini by 0.63, arXiv study finds, exposing an architecture-dependent split.
Technology & AI
FactorDiff routes diffusion experts by pixel on ARC-AGI
FactorDiff decomposes diffusion samples into pixel-level factors and routes each to the best expert, beating global weighting on ARC-AGI reasoning tasks.
Technology & AI
AMT-X multi-turn red teaming hits 100% attack success on frontier LLMs
AMT-X multi-turn red-teaming framework hit up to 100% attack success on six frontier AI models, exposing critical gaps in current LLM safety testing.
Technology & AI
PromptGraph turns LLM prompts into privacy graphs
PromptGraph models LLM prompts as graphs to catch privacy leaks that existing sanitisers miss, affecting anyone sending sensitive data to cloud AI.
Technology & AI
Best AI agents fail 50% of visual tool tasks, Apple benchmark shows
A new open benchmark with 500+ tools reveals even frontier AI models can't reliably read an image and act on it, with most failures traced to seeing, not