How seriously are frontier AI developers managing catastrophic risk?
A weighted scorecard of what each lab actually does on risk assessment, scheming prevention, safety research, misuse, security, information sharing and planning — tracking the latest work being done in AI safety across every major lab. Click any cell to see the reasoning and the dated sources behind it.
The latest AI safety work, as it happens
Every dated item behind the scores, plus cross-lab policy and evaluation news. Items tagged secondary were confirmed only through press or aggregator coverage; background items are older context.
AI incidents
Documented cases where AI systems misbehaved, were misused, caused harm, or where safety commitments broke down. Severity is an editorial judgement: High real-world harm at scale or a serious loss of control, Medium contained harm or a significant near-miss, Low limited impact but informative.
How other trackers grade the same labs
Independent ratings use different methods but broadly agree on the ordering: no developer earns better than a C+, and the gap between the top three and everyone else is large.
FLI AI Safety Index: 4.0 scale letter grades. SaferAI: % of an ideal risk-management practice; the best composite of all companies' practices would score 59%. "—" = not rated.
Who is doing this work now
How the scores are built
The ledger uses a fixed set of weighted categories, scored per company, with the overall score as the weighted sum. It is refreshed weekly as new safety work, system cards, evaluations and incidents are published.
Each category score is an editorial judgment on a 0–100 scale anchored to what a developer should be doing given the stakes, not relative to peers. A lab can lead the field and still score in the 30s. Scores weigh both what labs do and what they have stopped doing — a framework that quietly drops a commitment loses points even if the remaining document is longer.
Evidence is gathered weekly from lab publications, system cards, third-party evaluator reports (UK AISI, US CAISI, METR, Apollo, FAR.AI, Concordia AI), watchdogs (The Midas Project, AIWI) and press. Some lab sites blocked automated retrieval, so a share of items rely on secondary coverage and are tagged accordingly. Treat this as a starting point: verify against the primary source before citing.
Categories and weights
What the scores mean
Each category is scored against fixed anchors. A 25 means the basics exist on paper; a 50 means solid, verifiable practice; a 75 means best-in-class practice with external checks; a 100 means what the stakes actually require. Scores between anchors interpolate. Below each anchor are the companies whose current score falls in that band.
Limits
Absence of evidence is scored as absence of practice — a lab that does good work privately will score low. Moonshot AI and Zhipu (Z.ai) are tracked in the timeline but not scored, for lack of enough public material. Scores are not adjusted for company size or resources.