Blog
Notes and analysis on AI development.
Microsoft expands Copilot ISO 42001 scope to Studio agents
Microsoft is expanding Copilot ISO 42001 coverage to Copilot Studio, GitHub Copilot, Dragon Copilot, and Copilot Health.
AWS Nova Act Service Card defines the limits of browser agents
AWS documented Nova Act limits for browser agents, including 100 sequential steps, 30-minute sessions, prompt injection boundaries, and IAM resources.
QVAC AI SDK Provider lets local models plug into AI SDK apps
Tether QVAC published a Vercel AI SDK provider for local OpenAI-compatible servers. Here is what it changes for agents, TypeScript apps, and local AI routing.
Cohere Command A+ ships as an open-weight MoE for two H100s
Cohere Command A+ combines Apache 2.0 open weights, a 218B MoE design, 25B active parameters, 128K context, and enterprise deployment options.
SageMaker Adds OpenAI API Support for AWS-Hosted Models
AWS SageMaker now supports /openai/v1 endpoints, lowering the migration cost for OpenAI SDK, LangChain, Strands Agents, and AI gateways.
MiniMax M3 brings 1M context to open-weight coding models
MiniMax M3 combines 1M context, multimodality, and coding-agent benchmarks, but its weights and technical report are still pending verification.
Mistral Search Toolkit separates RAG failures from model failures
Mistral Search Toolkit public preview treats RAG and agent-search failures as retrieval, pipeline, and evaluation problems.
RTX Spark Debuts as a Local AI PC for 120B LLMs
NVIDIA and Microsoft introduced RTX Spark, a Windows PC category for local agents with 120B LLMs, 128GB unified memory, and OpenShell.
Codex Tax AI handled 7,000 returns, and the improvement loop starts with evals
OpenAI and Thrive showed how Tax AI links production traces, practitioner corrections, evals, and Codex tasks.
ChatGPT Sheets security report exposed prompt injection across sidebars
PromptArmor disclosed a ChatGPT for Google Sheets exfiltration path, and OpenAI removed Apps Script code generation.
Mythos found 10,000 vulnerabilities, now patching is the bottleneck
Anthropic Project Glasswing says Claude Mythos Preview and partners can find vulnerabilities faster than teams can validate, disclose, and patch them.
SkillOpt Turns Agent Skills Into Trainable Deployment Artifacts
Microsoft SkillOpt treats SKILL.md-style agent instructions as trainable artifacts updated through rollouts, validation scores, and bounded edits.