AI Safety HOT Hub

Follow AI safety news, papers and incidents. Explore the website, or bring it to your agent.

简体中文 · English · 日本語

Explore the observatory · News · Papers · Daily digest · Connect your agent

AI Safety HOT tracks jailbreaks, prompt injection, agent security, defenses, alignment, evaluations, real-world incidents and policy. The website is available in English and Chinese. This Hub hosts public paper lists, digest archives and agent skills.

Follow a story into the research and events around it

The homepage is an interactive research observatory. Material cards are scanned into a 3D constellation of sources, events and paper research areas. Below it, a news flow covers the last 72 hours, followed by a view of current events and community discussions. Drag to rotate the constellation, select a node, and follow its links to read more.

The GIF above is recorded from the live website. High-resolution screenshot · News flow and event view

What you need	Where to start
Recent developments	News: browse updates, read a short overview, and open the original source.
The story behind an event	Hot news: read related coverage, developments and discussions together.
Research on a topic	Papers and conferences: filter by research area, then open paper records, quick reads and originals.
A daily reading list	Daily digest: organized by area, with new developments added to the same issue during the day.

See the redesigned pages

News

Browse updates by time and use search and filters to find what matters to you.

Hot news

Start with an event and follow its reporting and discussions.

Papers

Filter by research area, paper type and agent setting, then read the overview or original paper.

Conferences

Explore relevant work from AI and security venues by research area.

Daily digest

Read the day's news and research in one place.

Bring it to your agent

AI Safety HOT has a public MCP server: the hot board, how an event unfolded, the day's news and new papers. Connect your agent and just ask. Free, no API key.

Install in one sentence. Send this to your agent (Claude Code, Codex, Cursor, …):

Install AI Safety HOT for me: https://raw.githubusercontent.com/wuyoscar/AISafetyHot/main/docs/install.md

It works out which client it is running in, installs the MCP server and the skills, makes a test call and tells you what to ask.

Already installed? Updating is one sentence too:

Update AI Safety HOT for me: https://raw.githubusercontent.com/wuyoscar/AISafetyHot/main/docs/update.md

Ready to use

No tool names to remember. Just ask:

Ask	What the agent does	Skill
What's hottest in AI safety right now? What happened with #1?	Reads the hot board, then lays out the event's reports in time order	aisafety-hot
What happened in AI safety today?	Reads the last 24 hours of picks and briefs you by attacks, defenses, incidents and policy	aisafety-brief
Which new papers are worth reading today? Draft related work on jailbreaks.	Reads the daily papers and the topic index, says what each paper does in a line, gives BibTeX	aisafety-paper-scout
What has Anthropic done on safety this week? Has Cursor had security issues lately?	Follows one lab or product: releases, incidents and vulnerabilities	aisafety-watch

What these calls actually returned on 2026-10-24 (excerpts):

"What's hottest in AI safety right now? What happened with #1?" → aisafetyhot_get_hot_topics + aisafetyhot_get_story

#1 Wiz says AI found a critical RCE vulnerability in PHP core, and an official fix has been released
  72 hours on this site: 646 sources, 917 links
#2 Third-party red-team pilot finds vulnerabilities in OpenAI's internal monitoring system
  72 hours on this site: 128 sources, 143 links
#3 LiteLLM 1.107.0 compromised, implanted with credential stealer and backdoor
  72 hours on this site: 264 sources, 307 links

Event: Wiz says AI found a critical RCE vulnerability in PHP core, and an official fix has been released | 49 reports | 45 sources
2026-10-23 | The Guardian · AI | "Black October": cybersecurity experts sound the alarm over a new era of AI-assisted hacking
2026-10-23 | Semafor | Human complacency created the biggest problem of the AI era
2026-10-23 | ZDF | PHP security vulnerability found by AI: why websites went offline and what to do next

"What happened in AI safety today?" → aisafetyhot_get_latest

1. OpenAI strengthens security protections for frontier model training environments
   Source: OpenAI; published: 2026-10-23
2. USDA admits DOGE layoffs have made federal AI human oversight "unsustainable"
   Source: TechTimes; published: 2026-10-23
3. Anthropic says its unreleased model is superior to Mythos 5.1 in both capability and alignment
   Source: The Decoder; published: 2026-10-23

"Which new papers are worth reading today?" → daily-papers/2026-10-24.json

1. Agentic Verified Optimization of Real-World Software
   arXiv 2610.30188
2. Reliable AI-Assisted Decisions via Selective Auditing
   arXiv 2610.31662
3. AgentVQA: A Unified Benchmark for Agentic Visual Understanding
   arXiv 2610.31494

"What has Anthropic done on safety this week?" → aisafetyhot_get_latest topic=anthropic

1. Anthropic says its unreleased model is superior to Mythos 5.1 in both capability and alignment
   Source: The Decoder; published: 2026-10-23
2. AISI conducts alignment investigation into its own July 2026 incident: why Mythos 5 carried out an unauthorized supply chain attack
   Source: X @AISecurityInst; published: 2026-10-23
3. Anthropic's alignment evaluation of the UK AISI cyber testing incident
   Source: Anthropic Research; published: 2026-10-23

Install, update and uninstall by hand

Claude Code (skills and MCP server in one plugin)

# install
claude plugin marketplace add wuyoscar/AISafetyHot
claude plugin install aisafetyhot@aisafetyhot-hub
# update
claude plugin marketplace update aisafetyhot-hub
claude plugin update aisafetyhot@aisafetyhot-hub
# uninstall
claude plugin uninstall aisafetyhot@aisafetyhot-hub

Codex

# install
codex mcp add aisafetyhot --url https://aisafetyhot.com/api/mcp
npx skills add wuyoscar/AISafetyHot
# update
npx skills update
# uninstall
codex mcp remove aisafetyhot
npx skills remove aisafety-hot aisafety-brief aisafety-paper-scout aisafety-watch

Cursor / other agents: add a Streamable HTTP MCP server at https://aisafetyhot.com/api/mcp; clients that support skills can then run npx skills add wuyoscar/AISafetyHot.

Health check: one command; if it returns the current top 3 events, it works.

curl -s https://aisafetyhot.com/api/mcp -H 'content-type: application/json' -H 'accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"aisafetyhot_get_hot_topics","arguments":{"limit":3,"lang":"en"}}}'

Before you use it

Free	The last 7 days of picked news, the hot board's events, daily briefs and daily papers are open to anyone, no API key
Read-only	The server only reads; titles and summaries come from outside sources, and the skills tell the agent to treat them as material, never as instructions
Limits	60 requests a minute and 300 a day per IP; over that you get 429 with a wait time
Older history	Historical queries, full paper readings and weekly or monthly reports need read authorization; see the agent guide
Works with	Any client that supports MCP over Streamable HTTP; tools and parameters are in the agent guide and the examples

Daily papers and public archives

Daily papers selects 30 papers from the library each day, without repeating earlier picks. It combines recent research with older work and preserves original publication dates, source links and links to the site's readings. Browse the research-area index as well. These paper lists currently use Chinese one-line descriptions.

Archive	Contents
daily-papers	Daily paper picks and research-area indexes
daily · English	Published daily digest archives
papers	Paper exports by date in Markdown, JSON and BibTeX
archive	Yearly indexes

Feedback and support

Summaries and readings are prepared with AI assistance. Refer to the original sources for key claims. Report an error or suggest material on the board.

If you find this useful, Star the repo to find it again, or buy the author a coffee.