Others doing autonomous or semi-autonomous AI-safety monitoring and briefing

Researched 2026-10-12. Most findings come from web search snippets plus a few page fetches; where I only saw a snippet it says so. The honest headline: I found no fully autonomous, safety-specific, actively maintained daily-brief agent with real adoption. Human-curated newsletters dominate, and the automated repos are small. Ria's differentiator (incident corpus + red-teaming + MCP) is genuinely open ground.

Top 3 to copy

OECD AI Incidents and Hazards Monitor - the only large automated incident pipeline: news clusters, LLM classification (incident / hazard / unrelated), embedding-based dedupe. Copy the pipeline shape.
Concordia AI's Frontier AI Risk Monitoring Platform - recurring, quantitative, quarterly report over 15 developers with a trend index. Copy "monitor a fixed panel on a fixed cadence and publish a trend," not just news.
AI Incident Database weekly snapshots + curation workflow - stable dated dumps, so Ria can diff week over week and cite incident IDs.

(Honorable mention for process: NayMyatMin/NewsUpdate as the nearest open-source analogue to what Ria's daily brief does; see below.)

Automated or semi-automated monitors

OECD AI Incidents and Hazards Monitor (AIM)

What: Global incident/hazard tracker built by the OECD.AI expert group on AI incidents, funded with support from the Patrick J. McGovern Foundation.
How: Event Registry (commercial news intelligence, >150,000 articles/day) supplies event clusters that contain AI concepts; LLMs classify events as incident, hazard, or unrelated. A secondary comparison site describes an embedding step: top-10 similar existing summaries by cosine similarity, then an LLM decides if the item joins a cluster (secondary source, treat with caution). OECD itself says coverage is partial and an open submission process is being developed.
Cadence: Continuous.
Copy: news-cluster -> LLM triage -> dedupe -> definitions of "incident" vs "hazard" (use the same two buckets in Ria so the catalogue maps onto OECD's).
URLs: https://oecd.ai/en/incidents-methodology , https://oecd.ai/en/eventregistry , https://oecd.ai/en/wonk/incidents-monitor-aim . A 2026 paper on classification: https://arxiv.org/pdf/2604.21412 (title: "A pragmatic classification framework for AI incident monitoring"; I saw only the listing, not the content).

Concordia AI (Beijing) - Frontier AI Risk Monitoring Platform and "AI Safety in China"

What: Platform launched with an inaugural 2025 Q3 monitoring report; tracks models from 15 developers across cyber offense, biological, chemical and loss-of-control risk, with a Risk Index. They report cumulative-max index growth of 31% (cyber), 38% (bio), 17% (chem), 50% (loss of control) over a year. Self-reported by Concordia; no independent methodology review found.
Cadence: Quarterly reports; the "AI Safety in China" Substack was at issue #19 (Jan-Mar 2025 coverage, delayed) when I looked; later issues not checked. Also a "State of AI Safety in China (2025)" report.
Copy: fixed panel of labs, fixed risk domains, dated index; plus a China-sources lane (Ria has no Chinese-language coverage today).
URLs: https://aisafetychina.substack.com/p/10-key-insights-from-concordia-ais , https://aisafetychina.substack.com/about , https://aisafetychina.substack.com/p/state-of-ai-safety-in-china-report

AI Incident Database (AIID)

What: Crowd/editor-curated incident catalogue. Monitoring is human-in-the-loop (submissions plus editors).
How/cadence: Weekly point-in-time snapshots (JSON, MongoDB archive, CSV; ~105-107 MB) generated by a GitHub backup workflow; newest seen 2026-08-10 on cached pages. No public API docs found in search.
Copy: snapshot diffing; stable IDs; the "contact us about your use" courtesy. Email them once Ria starts consuming the dump.
URLs: https://incidentdatabase.ai/research/snapshots/ , https://incidentdatabase.ai/research , https://github.com/responsible-ai-collaborative/aiid

Epoch AI

What: Data on AI (models, benchmarks, compute, data centers, hardware).
How: Curated databases published as CSV with CC-BY licensing; a data hub, a benchmarking hub, and an epochai Python client that reads Airtable (per search snippets; I did not confirm there is a REST API). A third-party page claiming data through 30 Sep 2026 looked low quality, so verify freshness on epoch.ai.
Cadence: Continuous updates; monthly-ish "Epoch AI Brief" posts historically.
Copy: every number in Ria's model-lineup resource should carry a source and date; consume Epoch CSVs instead of scraping.
URLs: https://epoch.ai/data , https://epoch.ai/blog/announcing-epoch-ai-data-hub , https://epoch.ai/benchmarks/use-this-data

AI Digest (theaidigest.org) and the AI Village

What: Founder Adam Binksmith. The AI Village puts frontier-model agents in a shared computer environment with goals (started April 2025 as a 30-day charity pilot with four agents; later grew, sources disagree on size).
How: Livestreamed agents; memory consolidation that rewrites notes over time.
Relevance: Closest public example of observing many agents interacting, which is what Ria's Moltbook red-teaming is. Not a briefing product.
Copy: public, replayable logs of agent behavior; memory-consolidation pattern for Ria's long-running notes.
URLs: https://theaidigest.org/village/blog/how-the-ai-village-works , https://theaidigest.org/village/blog/introducing-the-agent-village

AI Lab Watch (Zach Stein-Perlman)

Status: Per a SPAR project listing, dormant; SPAR is recruiting fall 2026 mentees to revive it (system cards, safety frameworks, a decision-maker dashboard). Successors such as the FLI AI Safety Index do similar reviews.
Copy: the per-lab scorecard over system cards and frameworks is a good Ria entity-profile schema. The revival project is a collaboration opportunity.
URLs: https://ailabwatch.substack.com , https://sparai.org/projects/f26/recrMjUJvwDNnEZlK

Open-source daily AI digest repos (2026)

Only snippets seen except where noted; none verified for sustained activity.

NayMyatMin/NewsUpdate - https://github.com/NayMyatMin/NewsUpdate (fetched). Daily AI-safety digest: fetches ~1,800 articles from 110+ sources (blogs, arXiv, Google News queries, WeChat via Sogou, RSSHub), filters stale/duplicate, LLM ranks and summarizes top 15 into four sections (Threats and Incidents; AI Security from Major Players; AI Agents and OS Integration; Research and Regulation) plus a weekly trend summary. gpt-4o-mini screens, gpt-4o analyzes; Anthropic and Ollama supported. GitHub Actions at 01:00 UTC. 0 stars, README says "Internal use," no OSS license: read for design, do not copy code.
CR7SMS/News-Digest - https://github.com/CR7SMS/News-Digest - scores articles on five axes, emails top 10, with a small eval framework. Daily 14:00 UTC (snippet).
hoangsonww/AI-News-Briefing - https://github.com/hoangsonww/AI-News-Briefing - multi-agent pipeline, 9 topics, delivers to Notion/Slack/Teams (snippet).
martinolszanecki-dot/margulis94 - https://github.com/martinolszanecki-dot/margulis94 - "so what" per item (snippet).
yeshuai508/ai-daily-news - https://github.com/yeshuai508/ai-daily-news - RSS + X, dedupe, LLM curate, optional podcast narration (snippet).
Copy: NewsUpdate's four-section structure, the cheap-model-screens-then-strong-model-analyzes cascade (cost win), and the five-axis scoring idea from News-Digest. Reuse the cascade in Ria's cost ledger.

Human-curated newsletters (the benchmark Ria is measured against)

Who	What / how	Cadence	Copy for Ria	URL
AI Safety Newsletter (Center for AI Safety)	Short issues on AI safety/policy; >23k subscribers per archive page.	Described as weekly, but 2026 archive shows #68-#73 from February to late May with multi-week gaps.	Brevity; "what happened / why it matters" blocks. Ria can fill the gaps between issues.	https://newsletter.safe.ai/archive
Transformer (Shakeel Hashim)	Weekly briefing on AI and AI policy, aimed at policymakers; also media coverage roundup.	Weekly	Policy-reader framing and sourcing; weekly roll-up on top of the daily brief.	https://transformernews.ai/about
Don't Worry About the Vase (Zvi Mowshowitz)	Very broad, opinionated weekly "AI #N" posts; one reviewer counted 11 posts and ~97k words in half a month. Process is not documented in sources I found.	Weekly+	Link-dense, per-item commentary, running scoreboard of claims. Note Ria cannot match judgement; aim for coverage Zvi can mine.	https://thezvi.substack.com (not fetched)
Import AI (Jack Clark)	Weekly; a few papers per issue in depth plus commentary; subscriber counts conflict (about 70k-90k+).	Weekly	Depth-over-breadth section for one paper per week.	https://importai.substack.com
Alignment Newsletter (Rohin Shah)	Started 2018, grew into a paid-summarizer team in 2019; last issue seen was #172 ("sorry for the long hiatus"); Shah's site describes it as on indefinite hiatus.	Was weekly	Cautionary tale: a research-summary service dies when the human summarizers do. Ria's value is automating the summarization layer.	https://www.alignmentforum.org/posts/rxsdSgnZrgYWc2XAp/an-172-sorry-for-the-long-hiatus
Epoch AI Brief	Monthly digest of Epoch work.	Monthly	Dated, sourced numbers.	https://epochai.substack.com

Academic work on automated horizon scanning for AI risk

My searches did not find a paper that describes itself as automated horizon scanning for AI risk. Nearest items, abstracts not read (only listings/references surfaced), so relevance is unconfirmed:

ForesightSafety-SAGE: automated scenario generation and safety evaluation for LLM agents - https://arxiv.org/pdf/2606.08531
AURA: agent autonomy risk assessment framework - https://arxiv.org/pdf/2510.15739
Risk analysis techniques for governed LLM-based multi-agent systems - https://arxiv.org/pdf/2508.05687
OECD-style classification: https://arxiv.org/pdf/2604.21412

Suggested follow-up: one targeted pass on Google Scholar / Semantic Scholar for "horizon scanning" + "AI" + "LLM" before claiming novelty.

Not covered / gaps

Galileo / Arize "AI news agents": searched only indirectly; no evidence found. Skipped rather than guessed.
Zvi's actual workflow and Import AI's tooling: no primary description found.
"AI Safety Papers" aggregators: not located in this pass.