Agents that write code increasingly also need to read research, and AI safety research moves fast. AISafetyHot-Hub, a repo with about 1,600 stars, packages a bilingual AI safety news and paper site as something an agent can call: a Claude Code plugin, six agent skills and an MCP endpoint that needs no API key.
Its README headline translates as "one command so your agent understands AI safety every day." This post covers what the repo contains, how to install it, and the open questions about sourcing that the README leaves unanswered. It is a tool write-up, so everything below is attributed to the repo and its companion site.
Quick reference
| Question | Answer |
|---|---|
| What is it? | AI safety news and paper monitor with agent integrations |
| Topics | Jailbreaks, prompt injection, agent security, evaluation, incidents, policy |
| Daily output | 30 curated papers, grouped by topic |
| Skills | Six: safety brief, paper scout, related-work drafting, lab watch, incident check, MCP tool reference |
| Install (Claude Code) | /plugin marketplace add wuyoscar/AISafetyHot-Hub then /plugin install aisafetyhot@aisafetyhot-hub |
| Other agents | npx skills add wuyoscar/AISafetyHot-Hub or MCP at https://aisafetyhot.com/api/mcp |
| API key | Not needed, per the README |
| Languages | READMEs in Chinese, English and Japanese; main README in Chinese |
| Activity | About 1.6k stars, 18 forks, 491 commits on main |
| License | LICENSE file present; the README page does not state the type |
What does the repo contain?
The README describes this layout:
skills/: six agent skills.daily-papers/anddaily/2026: the daily lists of 30 papers.papers/2026: paper archives; the site hosts about 2,000 AI safety papers.docs/: an agent integration guide..claude-plugin/,.mcp.json,AGENTS.md,status.json,archive/andassets/.
The six skills map to practical jobs. A safety brief summarizes the day. A paper scout finds relevant papers. A related-work drafting skill helps write the background section of a paper. A lab watch follows what labs publish. An incident check looks up reported incidents. And an MCP tool reference documents the endpoint's tools.
What does the site offer?
The aisafetyhot.com site is bilingual (Chinese and English). Its sections include an overview, hot news, live news, live papers, a safety daily, conference papers and starred news. It also has an email newsletter, a changelog and a message board. The overview shows a news stream for the last 72 hours on topic bands, and a hot-topic view that groups several sources about one event and shows how many people are discussing it. When we fetched the page it said 140 items were collected that day, without saying how many were papers.
If you only want to read, the site works without installing anything. The repo adds the agent-facing layer.
Which AI safety topics does it cover?
The README names six areas, and each maps to a different kind of reader.
| Topic | What papers in it usually study | Why a builder cares |
|---|---|---|
| Jailbreaks | Prompts or inputs that get a model to ignore its rules | Decides how much you trust a model's refusals |
| Prompt injection | Instructions hidden in content an agent reads | Core risk for any agent with tools |
| Agent security | Sandboxing, permissions, tool misuse | Determines what an agent can break |
| Evaluation | Benchmarks and red-team methods for safety | Tells you whether a claim is measured |
| Incidents | Reported failures and misuse | Source of real-world lessons |
| Policy | Rules, standards and disclosures | Affects what you may ship |
A compass needle aligned with a single pebble, illustrating AI safety alignment research topics
This table is explainx.ai's summary of the topic list, not text from the repo.
How is it different from reading arXiv yourself?
The value is curation and packaging, not new research. Instead of scanning a firehose of preprints, you get a capped list of 30 a day, grouped by topic, with one-line summaries, and the ability to ask an agent to pull it into a brief or a related-work draft. The trade-off is the one already noted: you inherit the curator's filter and the unknown quality of the summaries. For a field where the same topics keep recurring, a short grouped list is a time saver; for deep work, the original papers are still the source.
What are people asking?
Is it only in Chinese? No. The site is bilingual and the repo has English and Japanese READMEs, but the main README is in Chinese, so English readers should use the English README and the site's English view.
Is it affiliated with a lab? Nothing in the README or site we fetched says so. It presents as an independent project.
Does it cost anything? The README says the site is free and no API key is needed. It does not describe rate limits for the MCP endpoint.
Can I self-host it? The repo holds skills, daily lists and archives, but the README does not describe running the site or the MCP server yourself; the endpoint it documents is hosted at aisafetyhot.com.
How do I install it?
For Claude Code, from the README:
/plugin marketplace add wuyoscar/AISafetyHot-Hub
/plugin install aisafetyhot@aisafetyhot-hub
For other agents that support skills:
npx skills add wuyoscar/AISafetyHot-Hub
Or connect any MCP client to the Streamable HTTP endpoint https://aisafetyhot.com/api/mcp. If you are new to the pieces, our guides to MCP and agent skills explain how a server and a skill plug into an agent.
One thing to weigh before connecting: an MCP endpoint hosted by a third party sends whatever your agent asks it for to that service, and returns text your agent will read. That is normal for MCP, but it means the endpoint is part of your agent's trust boundary. Our explainer on indirect prompt injection and the write-up on MCP rug-pull and tool poisoning cover why returned text deserves scrutiny.
How are papers chosen?
The README says the daily list picks 30 papers each day, without repeats, and only from "well-known researchers, top labs and leading universities", mixing recent and older work. Each paper gets a one-line summary and many link to fuller write-ups.
Two things to know. First, "daily" does not mean "new today": the 2026-10-20 list included papers dated months earlier, such as August 2026. Second, the selection rule favors established names, which is a reasonable filter for quality but also a bias: strong work from unknown authors is less likely to surface.
What does it not tell you?
Read this section before you build a workflow on the feed.
The README and site leave real gaps, and they matter if you plan to rely on the feed:
- How news items are produced. The page does not say whether items are written by people, models, or a mix, or how summaries are reviewed.
- Error handling. The site invites users to report mistakes on its message board, which suggests errors can occur. There is no stated correction policy.
- Selection criteria for news. The 72-hour news stream has no described criteria.
- License type. A LICENSE file exists, but the README page does not name it.
- Who runs it. The site has an acknowledgments page, but we found no methodology or disclaimer text.
None of this is unusual for a young hobby project. It does mean you should treat summaries as pointers to the primary papers, not as a substitute for reading them, and check any claim you plan to cite.
Who is it useful for?
Researchers and students who want a daily reading list and a drafting helper for related work. Agent builders who want their agent to check the latest jailbreak or prompt-injection research before designing a defense. Security teams watching agent-security incidents; pair it with a source such as our AI agent security platforms overview. Journalists and policy staff who want a topic-grouped view of what is being discussed.
It is less useful if you need vetted, citable summaries or coverage of non-English-language policy beyond Chinese and English.
A sensible way to use it
- Start with the site to see whether the selection matches your interests.
- Install the plugin or skills in a sandboxed project, not a repo with secrets.
- Use the paper scout to find candidates, then open the original papers.
- Treat the incident check as a lead generator; confirm incidents against primary sources.
- Record which items you actually cite, with links to the paper, not to the feed.
Related safety coverage on explainx.ai includes China's AI labs and safety disclosures, and the REA reverse-engineering MCP server for agent-security tooling.
What to do if you run agents that read outside text
Feeds like this are useful precisely because agents read them, and that is also the risk: any text an agent ingests can carry instructions. AgentBeam, the agent security platform from the explainx.ai team, is built to stop AI agents before they take dangerous actions, which is the layer to add when an agent combines outside reading with tools that can change things.
Details reflect the repo README and site as of October 20, 2026. Star counts, install commands and the endpoint can change; check the repo.