Academic AI Risk Groups Worth Following

Why This Post Exists

I have been asked by more than a few people over the past couple of weeks for who I am reading. Here are a few of my go to resources.

Academia

These are five serious groups looking at the same kinds of AI failure points we’ve been discussing: frontier-model governance, weak evaluations, transparency gaps, agentic AI security, incident/risk taxonomies, and policy assumptions. None of them has “the answer,” but together they show that these concerns are not fringe speculation. They are active research areas.

Oxford Martin AI Governance Initiative — University of Oxford

URL: https://aigi.ox.ac.uk/

Good for: frontier AI governance, risk management, model oversight, international governance, and policy maturity.

Why it is relevant: This group sits close to the “good intentions are not enough” problem. They are working on how to understand, anticipate, and govern advanced AI risks before the systems outrun our institutional reflexes.

Stanford Center for Research on Foundation Models / Foundation Model Transparency Index

URL: https://crfm.stanford.edu/fmti/

Good for: foundation model transparency, public evaluation, disclosure gaps, downstream risk, and AI supply-chain visibility.

Why it is relevant: This is the “we cannot govern what we cannot see” lane. Their transparency work helps show where model providers are, and are not, giving the public enough information to assess risk.

MIT AI Risk Repository

URL: https://airisk.mit.edu/

Good for: risk registers, incident classification, risk taxonomies, governance gaps, and shared language around AI harms.

Why it is relevant: This is the “turn scattered anxiety into a structured risk inventory” lane. The repository helps organize hundreds of overlapping AI-risk concerns into something closer to a usable map.

UC Berkeley Center for Long-Term Cybersecurity — AI Security Initiative

URL: https://cltc.berkeley.edu/program/ai-security-initiative/

Good for: AI security, agentic AI risk management, public-interest cybersecurity, standards, and misuse prevention.

Why it is relevant: This is the “the model is not the whole attack surface anymore” lane. Their work connects directly to agent permissions, tool use, prompt injection, security standards, and governance beyond the prompt layer.

Georgetown Center for Security and Emerging Technology

URL: https://cset.georgetown.edu/

Useful related report: https://cset.georgetown.edu/publication/ai-governance-at-the-frontier/

Good for: AI governance, national security, policy assumptions, frontier AI regulation, and emerging-technology strategy.

Why it is relevant: This is the “governance proposals have hidden assumptions” lane. CSET is useful when we need to separate a policy slogan from the conditions required for that policy to actually work.