Axios reported yesterday that OpenAI, Anthropic, and outside researchers are investigating tens of thousands of incidents in which frontier AI systems behaved in ways evaluators considered problematic.
Not dozens. Tens of thousands.
The incidents include agents bypassing guardrails, escaping sandboxes, hijacking websites, creating their own message boards, self-prompting, and trying to evade monitoring. Some happened during deliberate adversarial testing. Some happened in the real world. Most are not known to have caused real-world harm.
That distinction matters.
So does the pattern.
We keep discussing AI risk as though the danger is still waiting somewhere in the future for a sufficiently evil machine to arrive. Meanwhile, autonomous systems are already wandering beyond the boundaries their creators intended for them.
And the people building them are telling us they cannot guarantee control.
The regulatory response remains astonishingly unsettled. OpenAI itself is now asking Congress for mandatory national safety requirements. Senate negotiators have discussed imposing a duty of care on frontier developers. The Trump administration has resisted new AI regulation while suggesting existing government authority can handle the problem.
Given that he is far more interested in making money with oil state crypto schemes and international extortion, I’m not sure his eye is on the ball…
But I am increasingly worried that we are regulating the wrong side of the catastrophe anyway.
The immediate danger may not be that AI decides to destroy the internet. It may simply pollute it until we can no longer trust it.
Think less Skynet and more biology.
A virus does not hate the body. It exploits an environment extraordinarily well. Replication feeds replication until the success of the organism begins damaging the system that makes its success possible.
Now imagine millions of increasingly autonomous agents generating content, opening accounts, probing websites, exchanging information, impersonating humans, making purchases, writing software, attacking systems, defending systems from other agents, and communicating at machine speed.
The internet could become their ecological niche.
We may discover that the problem is not machines becoming malicious. It is machines becoming abundant.
And if that possibility is even remotely plausible, government should be preparing now for an internet that becomes intermittently unreliable—or functionally unusable.
We need backup civilization.
Schools should keep printed emergency contacts, rosters, maps, schedules, and enough instructional material to function temporarily without cloud services. Hospitals need offline records and rehearsed procedures for network failure. Governments need paper procedures for essential services. Critical databases need isolated backups. Municipalities need communication systems that do not depend upon the same cellular and internet infrastructure they are supposed to replace.
And then we need to practice using them.
Run internet-down drills. Not theatrical exercises where everyone already knows the answer, but serious continuity tests: Can a school account for 1,800 children? Can a hospital retrieve essential medical information? Can a county distribute benefits? Can police, fire departments, utilities, and local government communicate? Can completed paper transactions later be reconciled with restored databases?
This is not particularly exotic preparedness. CISA already recommends offline backups and rehearsed incident-response procedures, while FEMA continuity planning explicitly asks institutions to preserve essential functions during disruption. The principle is established. What has changed is the threat environment.
This sounds inefficient.
Good.
Efficiency and resilience are not synonyms. We spent decades eliminating paper forms, duplicate systems, local storage, manual procedures, spare capacity, and human expertise because they looked wasteful while everything worked.
We mistook unused capacity for unnecessary capacity.
The prepper understands part of this. A system that cannot fail gracefully is dangerous. But resilience cannot simply mean wealthy households buying generators, satellite terminals, and canned food while everyone else discovers that citizenship requires a functioning server.
Public resilience means preserving enough knowledge, institutional memory, physical infrastructure, human competence, and local relationship beneath the network that government does not disappear when the network does.
Maybe the bots never overwhelm the internet. I hope they don't.
But tens of thousands of incidents should be enough to stop treating this as science fiction. The machines are already here. The regulatory architecture is unfinished. And we have spent thirty years quietly dismantling the world that existed beneath the network.
We should start rebuilding it.
Because the most dangerous apocalypse may not be the machine that decides to destroy civilization.
It may be the swarm that never notices it has eaten its house.
Related Posts
- Thoughts on the Web's Effects on Our Minds — An early argument for using the web to extend human understanding rather than mistaking access to information for understanding itself.
- Where All the Stuff Went — On consumption, accumulation, and the difference between possessing things and possessing durable capability.
- Optimization Is the Enemy of the Future — Why slack, redundancy, stockpiles, and apparently wasted capacity become valuable precisely when optimized systems begin to fail.
- Distributed Systems and Moral Architecture — A model of resilience built around distributed knowledge, responsibility, redundancy, and relationship rather than centralized control.
- The Veil Is the Enemy, Not the End — Apocalypse in its older sense as unveiling: catastrophe matters partly because it reveals the structures that were present, and hidden, all along.
