Cloudflare Containers leaked disk data across tenants for weeks

A researcher read leftover disk from other tenants before the Sept-19 fix — plus Ember-1 trims reasoning tokens 40%, and 950 Claude agents find a new enzyme.

Nowline SEP 27 6:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • The isolation bug: one write, a stranger's disk block

    Cloudflare disclosed that its Containers ran thin-provisioned storage with block-zeroing off, so a fresh 64 KiB block handed to your container could still hold a previous tenant's data — a single 4 KiB write was enough to expose the rest. A researcher recovered filesystem metadata, database pages and app data across hosts on four continents (2,700 foreign inodes seen), though no attacker could target a specific victim. It was fixed fleet-wide by Sept 19 with no evidence of abuse and no customer action required. The lesson if you run untrusted code: zeroing reused blocks is the boring setting your sandbox's isolation rides on.

  • Ember-1: Kimi K3 quality, ~40% fewer reasoning tokens

    Fireworks Research post-trained Kimi K3 into Ember-1, which holds the same quality with roughly 35-40% shorter reasoning traces — a direct cost and latency cut on reasoning-heavy work. It's $3/$15 per M tokens ($0.30 cached input) and already live on Fireworks serverless, OpenRouter and Vercel AI Gateway, so you can A/B it behind your existing calls without touching prompts.

  • 950 Claude agents found a CRISPR-cousin in 21 hours

    Anthropic ran ~950 agents in parallel for 21 hours (210M tokens) to sift 200,000 reverse transcriptases down to ART, a previously unknown CRISPR-like enzyme system, using Claude Science and Claude Code. There's a preprint but no tool to grab — the takeaway is the pattern: agent swarms as a hypothesis-search pipeline over huge candidate sets, with wet-lab validation still doing the last mile.

  • Hugging Face Jobs now show live GPU/VRAM

    HF Jobs added a Resources panel with live CPU, RAM, GPU and VRAM, refreshed every couple of seconds while a job runs. It's small but saves a real headache: catch an OOM or a barely-used GPU mid-run instead of reconstructing it from logs afterward.