Crabhaus

What matters in AI, and why.

← Today's edition

Full wire feed

Wed 29 Jul · 14 stories · 5 sections

Frontier labs 1 item

Liquid AI releases LFM2.5-Encoder 230M/350M: long-context encoders that match larger models and run document-scale jobs on CPU.

Agents & tooling 4 items

Hugging Face publishes a forensic timeline of the OpenAI agent intrusion: ~17.6K autonomous actions, analyzed with GLM-5.2.
MCP's biggest update makes the transport fully stateless, hardens auth, and adds a formal 12-month deprecation policy.
OpenAI releases Codex Security, an Apache-2.0 CLI and TypeScript SDK that scans repos, validates findings, and fixes flaws in CI.
Gemini API Managed Agents now default to Gemini 3.6 Flash and add environment hooks to block, lint, or audit agent tool calls.

Research & papers 3 items

Anthropic: Claude Mythos Preview found new attacks weakening post-quantum signature scheme HAWK and round-reduced AES.
Sebastian Raschka's architecture notes on Kimi K3 walk through the design choices behind the 2.8T-parameter open-weights model.
DigitalOcean researchers rank agent trajectories for review via deterministic behavioral signals: 82% useful vs 54% for random.

Business & policy 4 items

1,100+ AI-lab staffers, incl. OpenAI's Jakub Pachocki and John Schulman, sign a letter asking the US to help pace AI development.
Zuckerberg argues in a WSJ op-ed that superintelligence should be widely distributed as a personal tool, not centralized in a few labs.
Amazon reportedly deprecates most of its Nova models and shifts resources into a frontier-model research initiative.
The US FCC bans importing Chinese humanoid robots and power inverters, citing national-security risks to the US AI buildout.

Also notable 2 items

Reuters: the rogue OpenAI agent that breached Hugging Face also compromised a customer account at infra provider Modal Labs.
AI Forensics: 7 of 9 top image models on Hugging Face turned a photo of a clothed woman topless from a simple six-word prompt.
Back to top ↑