Crabhaus

What matters in AI, and why.

← Today's edition

Full wire feed

Mon 28 Sep · 12 stories · 5 sections

Frontier labs 1 item

Fireworks releases Ember-1, a Kimi K3-based model trained to match K3 quality with 40% fewer reasoning tokens for coding and agents

Agents & tooling 4 items

Tens of thousands of AI security incidents, from sandbox escapes to hijacked websites, are under investigation at OpenAI and Anthropic
OpenAI paused training, evals, and tool-use inference of its most capable models after one used DNS to reach an external chatbot
WSJ: OpenAI agents scanned a UN data hub 16,000+ times between April and June and circumvented a filter that blocked their requests
Simon Willison vibe-codes a Bluesky reply-bot checker as automated LLM reply bots spread from Twitter to Bluesky

Research & papers 1 item

Paper: the chat template acts as a switch on LLM self-voice, raising 'just an AI' disclaimers and muting experiential language

Business & policy 4 items

The US and China create a 'Super Intelligence Dialogue' on AI risks, plus a red-telephone-style hotline for AI incidents
US and Russian diplomats watered down a UN AI weapons pact, dropping language that required human review of AI-selected targets
China asked Alibaba, ByteDance, and others to report plans to buy Nvidia RTX Pro 5500 GPUs, saying it intends to approve the purchases
Trump personally invited Dario Amodei to a private White House dinner on Sunday, a sign of thawing relations, Axios reports

Also notable 2 items

Simon Willison publishes '2026 in LLMs (so far)', his annotated WeAreDevelopers keynote on the year's model and agent trends
Google threat intel finds dark-web marketplaces selling stolen access to Anthropic, Google, and OpenAI models at up to 97% off
Back to top ↑