The AI CEOs Told the UN Security Council It's Time to Worry. What Your Team Should Do Now.

If you've been putting off writing down how your team is actually supposed to use AI, this week is a bad week to keep stalling. On September 23, Sam Altman and Dario Amodei briefed the UN Security Council on AI risk — the first time either CEO has addressed that body in person. Altman said "no level of catastrophic risk from AI is acceptable." Amodei went further, proposing a ban on AI-enabled bioweapons research, shared evaluation systems between labs, and incident notification standards that would make AI failures reportable events rather than quiet internal fixes.

What happened

Altman told the Council that "the most important decisions cannot be made by labs in San Francisco alone" and called for global safety benchmarks. Amodei's incident notification proposal is the detail worth sitting with: he wants AI failures treated the way airline near-misses or hospital adverse events are treated, as incidents that get logged, classified by severity, and shared, not fixed quietly and forgotten. Neither man was vague about timing. Both treated this as current, not theoretical.

The skill implication

Two specific proposals point at two specific gaps worth checking against your own team. First, incident notification assumes you can tell the difference between a low-stakes AI mistake and one that needs to be escalated and logged — which only works if someone has written down what counts as which. Second, shared evaluation standards assume your organization has something to compare against, which means having your own written policy on what AI is and isn't allowed to touch before a regulator or a client asks for it. If international bodies are moving toward auditable AI failure trails, the organizations that already have a policy and a severity scale are the ones that answer that question in a day instead of a quarter.

Two ways this plays out

A compliance manager at a 90-person regional bank uses an AI tool to draft first-pass responses to regulatory inquiries. Before this year, "review before sending" was the entire policy — true in spirit, useless in practice, because nobody had defined what a bad draft looked like versus a merely rough one. She now keeps a one-page severity scale: a draft that cites the wrong regulation is logged and escalated to her compliance lead within the hour; a draft that's just clunky gets rewritten and moves on. The scale took an afternoon to write and turned "someone should probably look at this" into an actual decision she can explain to an examiner.

A marketing director at a 15-person agency has no compliance department and no bioweapons-adjacent risk, and assumed none of this applied to her. But her team uses AI to draft client-facing copy, and one hallucinated statistic in a published client report became exactly the kind of incident Amodei was describing, just at agency scale instead of nation-state scale. She wrote a half-page policy after that: any AI-drafted claim involving a number, a date, or a named source gets checked against a source before it ships, no exceptions, logged in a shared doc if it's caught after the fact rather than before. Six months in, the doc has three entries. All three caught before a client saw them.

The takeaway

Governments are moving toward treating AI incidents as auditable events because waiting for someone else to catch the failure has already gone wrong often enough to be a UN Security Council agenda item. Writing your own severity scale and policy now means you're answering that question on your terms, not scrambling to produce one when a client or regulator asks for it first.