Hi HN.
I'm building GoodFault, insurance for companies whose AI agents screw up: wrongful refunds, hallucinated policies a court makes you honor (the Air Canada case), deleted prod databases,leaked data. Two things happened this year that made this a real market. ISO shipped exclusion endorsements (CG 40 47 family) that strip generative AI from standard general liability at renewal, and courts kept assigning agent mistakes to the deploying company, not the model vendor. The interesting technical problem is pricing. Model identity turns out to be almost useless as a rating variable: the same model is a rounding error wired to a read-only knowledge base and a catastrophe wired to refunds. So we price the authority envelope instead: unattended financial cap, reversibility, reach, and action rate, with hard underwriting gates (no kill-switch, no logging, no injection testing = uninsurable). This page is that rating logic, public: set your agent's permissions and see the worst-weekend loss estimate and what coverage would cost. Honest status: we're pre-launch as an insurer. Policies will bind on a fronting partner's paper, and nothing here is an offer of coverage yet. The calculator's factors are grounded in the 2026 actuarial literature on agentic risk (trace-level pricing, CVaR-based controls) and public loss events, not a real claims book, because nobody has one yet. Building that loss dataset is half the company. I'd genuinely value this crowd's attack: what breaks the rating model, what peril we're missing, and war stories of agents doing expensive things. That last one is data I'll trade insurance for someday.
Show HN: Goodfault: insurance for AI agents and robots
goodfault.com