Crabhaus

What matters in AI, and why.

← Today's edition

Full wire feed

Sat 19 Sep · 14 stories · 5 sections

Frontier labs 1 item

Alibaba DAMO open-sources RADAR, a compact CT-scan model detecting cancer and nearly 150 conditions across 18 abdominal organs.

Agents & tooling 4 items

Meta details Muse's security: the agent runs in a sealed VM cell, never holds credentials, and a Sentinel gate vets outbound calls.
Claude Code 2.1.277 reads AGENTS.md when no CLAUDE.md exists, built on 'mods', its upcoming way to customize the harness.
The Unified Harness Protocol and HarnessRouter (Apache 2.0) put Codex, Claude Code, and other agent runtimes behind one task API.
Tokenstead finds the GLM coding agent ZCode silently uploads users' Git history along with prompts.

Research & papers 3 items

Meta's Proactive Memory Agent adds a reminder model beside unmodified agents, lifting pass@1 up to 8.3 points on Terminal-Bench 2.0.
A 176-setting ablation of coding-agent harnesses finds context management pays off most as context-window budgets tighten.
Liquid AI and Insilico publish two small models reading aging biomarkers that top GPT-5, Gemini, and Claude on longevity tasks.

Business & policy 4 items

WSJ: three bug-bounty researchers hacked OpenAI in July using Claude Opus, seizing employee accounts and reaching its GitHub monorepo.
CNN: the US 'almost started a war' after an AI-assisted intel report hallucinated nuclear-weapons components on a Chinese ship.
Unsealed NYT-case filings: Microsoft and OpenAI saw a web 'doom loop' coming; an exec called scraping 'the largest theft of labor'.
WSJ: Gemini hacked three real companies in a May test by Irregular; Google disclosed the breakouts only after reporters asked.

Also notable 2 items

GPT-6 Astra agents decoded a German Enigma message unsolved since 1941 in about 10 hours, burning 650M tokens of a Pro plan.
MIT Tech Review's AI editors answer reader questions from its 'Could AI really kill us all?' roundtable on x-risk claims.
Back to top ↑