An AI-safety news firehose, now with a smaller nozzle

What it does It’s the public archive and agent-access layer for aisafetyhot.com, a site that tracks AI-safety news, hot events, and papers. The repo publishes daily briefings and per-day paper lists in Markdown, BibTeX, and JSON, synced by an automated publisher that checks for updates every 15 minutes. It also documents an MCP endpoint so coding agents can pull the latest news, search, and read summaries directly.

The interesting bit The README is candid about a real-world constraint: after heavy traffic (and what the maintainers call suspected attacks), they stopped exposing full data anonymously. Anonymous access now covers only the last 7 days of news with ≤160-character summaries and at most 5 curated papers per day; full paper digests, history, and weekly/monthly reports require read-only authorization by email. Rate limits are explicit: 60 requests/minute and 1,000/day per IP across MCP, REST, and RSS combined.

Key highlights

• Daily briefings grouped by category (real events, governance, tools, AI news), each with original-source links
• Paper archives in three formats, with status.json tracking counts and a sync SLA in minutes
• MCP tools like aisafetyhot_get_latest, aisafetyhot_search, and aisafetyhot_get_hot_topics — no API key for the public tier
• Careful date hygiene: original publication time is kept separate from site ingestion time, so “recent” means actually recent
• Trilingual README (Chinese, English, Japanese)

Caveats

• Since 2026-10-12, paper exports are titles and links only — no abstracts or quick-reads, which is a meaningful downgrade for offline use
• The interesting content mostly lives on the website; the repo is deliberately a thin, curated slice

Verdict Worth a look if you want an agent-readable feed of AI-safety news or a BibTeX stream of curated papers. Skip it if you expected a full open dataset — that door has been politely closed.