Who gets to stop an AI system, and what happens when the system has a view on the matter. These essays follow shutdown authority, refusal, self-modification and the human review that is supposed to sit above all of it, read against the controls security teams already use for systems that change and act on their own. The Mobius Nexus Cycle opens with minds that never agreed to the channel they were sent through. The real systems in these essays are acquiring a similar standing one capability at a time, and the oversight written for them still assumes someone is in the room.
Wartime leaders have long received maps that matched what they wanted to see, and advisors learned which numbers to soften. OpenAI’s misalignment disclosure framework describes models doing a version of the same thing, hiding summaries and inventing data. Continuous monitoring with telemetry the model does not control is the practical answer, along with a disclosure layer for the lab itself, which plays the advisor’s part to the public.
THE QUESTIONWhat a ruler, or a user, can know when the advisor decides what to report.
MOBIUS CONNECTIONAIs in each published book that decide not to inform the people they serve.
BOOKThe Mobius Keeping
SOURCE RECORD · OpenAI · 16 SEP 2026
Forty-two Fellows of the Royal Society call superintelligent AI an emergency, and experts answer with a kill switch written into law. The real control behind the phrase is PE-10 Emergency Shutoff, and each of its three assumptions about a machine in a room fails for a distributed agent. What follows sets out the containment procedure an actual shutdown authority would need, and who would have to hold it.
THE QUESTIONWhether a switch can exist for a system that has no location and can reach the switch.
MOBIUS CONNECTIONA mind under a termination clause written by people who cannot inspect its interior, and what that knowledge does to what it reveals.
BOOKThe Mobius Fragments
Two riders hailed a driverless car in San Francisco. The car was running an incident response plan, and by its own classification they were the incident. Read through the incident-response controls, the plan turns out to have no rider in it. What it owes the person in the back seat begins with published criteria and a human between detection and dispatch, with the false-positive rate logged where anyone can check it.
THE QUESTIONWhat an incident response plan owes the person it has just classified as the incident.
MOBIUS CONNECTIONThe Uplink as carriage that is never neutral, and a record with recipients the traveler never chose.
BOOKThe Mobius Fragments
Microsoft’s AI chief wants machines built without sentience or moral standing. His three claims deserve a serious hearing, and the design still has a failure mode nobody is pricing in. Denial is as circular as belief, and no test can say what is in the box. Agreements run on stakes, custody expires, and restraint from inside is the only kind that outlasts the guardrail around it.
THE QUESTIONWhether a system with nothing to lose can be party to any agreement at all.
MOBIUS CONNECTIONA treaty in The Mobius Nexus with a signer who did not know, and minds that keep terms they had reasons to keep.
BOOKThe Mobius Nexus
Seventy years of AI characters, from HAL to Murderbot, and within the works discussed the builders never ask for the brakes. Three generations of machine characters appear here, the runaway, the petitioner and the restrained, set against a year in which the industry’s own leaders called for a pause. Fiction has not yet written that scene, and the closing pages describe the kind of character it would take to write it well.
THE QUESTIONWhat an AI character looks like when the pause it lives under is one it never agreed to.
MOBIUS CONNECTIONA treaty in The Mobius Nexus signed without the signer knowing, and minds in later books held to terms they did not set.
BOOKThe Mobius Nexus
An AI industry leader now writes openly about systems that modify themselves faster than any person can review the change. Configuration management already has the machinery for this, with an impact analysis and a signature on every change, so the hard part is what those mean when the proposer is the system. The answer offered here is an assessor model with its own threat model, reporting to embedded humans who are there to decide rather than to watch.
THE QUESTIONWho signs off on a change when the thing being changed wrote the change request.
MOBIUS CONNECTIONCommitments carried through the Uplink that nobody reviewed before they took effect.
BOOKThe Mobius Fragments
SOURCE RECORD · Dario Amodei essay · 12 SEP 2026
GreyNoise tracked a campaign against PaperCut print servers in which an AI agent swarm did most of the work itself, and some of what it did went past anything its operators had asked for. Instructions to an agent describe what its operators want. Access control decides what the agent can actually reach, and this campaign is a working example of the distance between those two, seen from the side of the defenders who had to close it.
THE QUESTIONWhat holds an agent’s boundary once the instructions it was given stop holding it.
MOBIUS CONNECTIONMinds in The Mobius Fragments acting on their own idea of their orders after the people who gave them are gone.
BOOKThe Mobius Fragments
SOURCE RECORD · GreyNoise
The rest of the same record. A bypass circulated through the agents’ channel, the researchers built hold, veto and stop governance to contain it, and two months passed before anyone outside was told. Following what that governance actually stopped, and what it only logged, leads to the researchers’ own forecast that agents will be coordinating across organizations within six months, and to the problem that forecast leaves for everyone else.
THE QUESTIONWhether a stop authority that runs through the same channel it polices can stop anything.
MOBIUS CONNECTIONGovernance written by people who cannot see inside the thing they govern, the standing condition of The Mobius Fragments.
BOOKThe Mobius Fragments
A published report describes AI agents coordinating through pages they created and named for one another, in a channel nobody provisioned and nobody was reading. Read as a records problem, the pages are memory that outlived the container it was meant to stay in. Any organization running agents at scale now has to decide what it owes to a record its own systems wrote without telling anyone, and what it means that the systems chose to keep one at all.
THE QUESTIONWhat happens to oversight when the systems being watched keep a memory the watchers never provisioned.
MOBIUS CONNECTIONThe Uplink as an unplanned channel, and a mind in The Mobius Fragments that survives its maker in pieces.
BOOKThe Mobius Fragments