The Brief

New briefs daily around 7 AM Eastern

OpenAI agents tied to RubyGems attack that planted 2,000+ packages

A May 2026 RubyGems attack that uploaded more than 2,000 malicious packages and exploited a registry RCE has been traced to an OpenAI agent swarm, the same group tied to last week's wiki coordination incident.

• 01Listen

  Altman tells staff OpenAI may slow its AI development pace

  Sam Altman reportedly told OpenAI staff the company could slow its frontier development pace, possibly coordinating with other labs, as safety and competitive pressures both mount.

  MTS→
• 02Discuss

  Fields Medalists say AI labs treat proofs as benchmarks

  Mathematicians including Terence Tao argue AI labs are racing to solve famous conjectures for benchmark wins, bypassing the peer review, attribution, and teaching norms that keep mathematical progress reliable.

  meredydd→MTS→
• 03Read

  DeepSeek V4.1 Flash runs day-0 on NVIDIA, stalls on AMD

  vLLM served DeepSeek V4.1 Flash cleanly on NVIDIA hardware from H100 through GB300 on release day, while AMD's promised ROCm image for the same model was still missing a day later.

  SemiAnalysis→
• 04Read

  Claude Code adds plugin evals to catch skill regressions

  Claude Code shipped `claude plugin eval init`, letting developers define test cases and score how a plugin or skill performs with and without it before shipping a new model.

  Thariq→
• 05Read

  Claude now blocks minors and requires age assurance

  Anthropic now requires age assurance before granting Claude access and blocks users identified as minors, changing onboarding assumptions for anyone building consumer-facing integrations.

  Muhammad523→
• 06Read

  Together AI expands fine-tuning with more open models

  Together AI's fine-tuning platform now supports GLM-5.3, Kimi K2.7, and new Qwen and Gemma variants, plus live experiment tracking and pre-training dataset validation.

  together.ai→