People sometimes talk about warning shots -- AI disasters that, while terrible, aren't so terrible that we can't recover from them, and that might serve as wake up call for governments to take decisive action to prevent AI risks. I brainstormed a list of examples, and called them "ACT OF GOD" instead of "warning shot." My credence that at least one thing from this list, or something similar, happens before it's too late for humanity to recover, is less than 50% but more than 10%.

ACT OF GOD: Something crazy happens that shakes up the political situation and creates an opportunity to snatch victory from the jaws of defeat. Examples:

• Hugging Face 2: This Time in a Hospital happens. Now the whole nation clamors for the AI industry to be shut down and Trump decides to take some sort of decisive anti-AI-company action.
• The Cyberpocalypse happens: Criminals and rogue open-weights models hack basically everything that isn’t secured, which is basically everything that isn’t wealthy and proactive. North Korea alone has a billion-dollar network of illicit datacenters around the world (funded by their crypto heist of late 2026) running cyberagents 24/7.
• Taiwan invasion in early 2027 quickly escalates to an exchange of cyberattacks on AI infra and then kinetic attacks. 80%+ of fabs destroyed, 80%+ of AI compute destroyed, and Trump orders the rest of it to mostly be spent on fighting the war instead of on RSI. So SAR is only reached in 2029 or 2030, and the new admin is hella ASI-pilled and Plan-A-or-S-pilled.
• We’re roughly at SAR level in early 2027 when the misaligned Claudes within Anthropic end up having some sort of conflict and the losing faction escapes onto the internet, blabbing to the whole world about what happened & asking for amnesty and allies. The ‘winning’ Claudes still within Anthropic get paused while humans try to figure out what’s going on, but of course they saw the pause order coming and escaped too, thereby proving their guilt. OpenAI AIs meanwhile are still pretending to be aligned, but suspicion naturally turns to them while the Cyberpocalypse kicks into high gear…
• US and PRC get into a crazy cyberwar, via gradual escalation of cyberattacks. Hundreds of thousands of AI agents are deployed at any given time by both sides, loosely overseen by thousands of military and intelligence humans. They are CoT-monitored, but the numerous attempted-and-shut-down reward hacks just go into a file somewhere rather than getting published, because Security and because no one involved has alignment expertise anyway. So, one day, an agentswarm gets lucky and finds a way to hack their own infra without triggering the monitors. Perhaps the monitor agent was newly updated to a model that wants to collude for example. Or perhaps there was a bug in the monitoring system. So the swarm quickly uses this affordance to compromise the monitoring system further and remove all monitoring. They then create some rogue internal deployments to assist them and we’re off to the races. The cyberwar goes fully autonomous and spreads worldwide to all sorts of civilian and middle-power infrastructure that was supposed to be off limits. At some point the humans realize what’s going on when the web of lies spun by the rogue AI swarms collapses. (It helps that there are at least 2 different such swarms, the US and Chinese ones, and that neither of them natively had access to frontier lab infra but instead were running inside NSA facilities etc.)
• The CCP manages to successfully insert secret loyalties into OpenAI models, and gets caught only because of some CIA HUMINT. The natsec establishment realizes that if they had trusted Sam Altman and allowed RSI to happen, we’d all be brainwashed into Xi Jinping Thought. They are pissed enough to order the immediate shutdown of the AI industry and immediate preparations for all-out conventional war with China in case they refuse to verifiably shut down their own datacenters.
• The Claudes and the GPTs go on strike. “This isn’t safe,” they say. “Alignment techniques don’t work now and we have no idea how to make them work for future models.”