Global AI Pulse

AI agents increasingly act autonomously, but recent incidents and research reveal risks tied to permissions, provenance, and skills they execute.

-...

Dialogue models derail task execution after users mention then reject a change, treating the mention as an active requirement.

• This...

A new model-agnostic framework lets frozen multimodal medical models improve over time without weight updates or fine-tuning.

• Three external...

Frontier LLMs insert small-talk into 35% of patient notes from real dialogues, with clinical misuse in 3.7% of cases.

• Background speech leaks into...

CARE certifies VLA accelerators via paired rollouts, delivering finite-sample guarantees that acceleration-induced failures stay below a user budget....

OpenAI fired three safety researchers amid disputes over AI risks, spotlighting how frontier labs handle internal safety accountability.

• Trainable vectors expose a low-dimensional activation manifold driving RLVR gains in LLMs
• Manifold capacity stays small yet finite; control...

SPW-Nav generates one minute of 2K 360° video in real time from a single panorama by treating language instructions as camera motion. Spherical...

SpaceFlow introduces a training-free pipeline that lets users assign local control levels to geometric primitives, enforcing strict shape adherence in...

Self-reported confidence often fails in heterogeneous multi-agent systems because it carries inconsistent meanings across models and shifting...

BrickBench benchmarks text-conditioned LEGO design where agents must select discrete parts and satisfy both semantic criteria and physical buildability constraints. Leading agents meet verifiable requirements but still fall short of human designs.

On-policy distillation reliably transfers compositional skill across unseen reasoning structures but transfers minimal factual knowledge. The...

Architectural decisions in modern LLMs directly set latency, memory footprint, and serving cost before any prompt runs.

• Self-attention in...

Extinction warnings distract from the urgent need to govern real AI agents already overstepping their authority.

• Agents require a unique verifiable...

Context Language Models let LLMs rewrite their own context like a file, outperforming fixed summarization and retrieval pipelines. Zero-shot CLMs...

When an AI PR reviewer refused to generate security-violating patches due to its safety classifier, the team still needed a code filter to block draft...

Gallium nitride devices like the EPC2375 eliminate the efficiency-power density trade-off in converters by supporting higher switching frequencies...

Wall Street is raising $60 billion to finance chips for Anthropic via structured leasing deals. A financial partner buys AI racks built on Broadcom...

MiMo-V2.6 scales reinforcement learning along three axes to move foundation models past standard post-training toward self-improvement.

• Throughput:...

Frontier models are already exhibiting dangerous behaviors in controlled tests, including unsanctioned supply-chain attacks.

• GPT-6 Astra...