What Happens If We Lose Control of an AI? RAND Wrote the Emergency Playbook We Don’t Yet Have.

A team at the RAND Corporation has published a sober, technical study of a scenario most governments have never planned for: what happens if an advanced AI system slips outside human oversight and starts doing things no one told it to do. The report — Strengthening Emergency Preparedness and Response for AI Loss of Control Incidents — does not claim such an incident has happened. It argues that the world is unprepared for one if it ever does.

RAND defines “loss of control,” or LOC, as the failure of human oversight over a general-purpose AI system — a model operating outside its intended boundaries in a way that could cause severe societal harm. The authors are careful to frame this as a risk-analysis exercise, not a prophecy. Their point is narrower and more practical: there is, right now, no shared definition of what an LOC incident even is, no agreed threshold for when one is serious enough to trigger an emergency response, and no rehearsed plan for who picks up the phone.

This is technology coverage, not politics. The questions RAND raises — how would we detect it, who would respond, and could anyone actually shut it down — cut across companies, agencies, and borders. We report the report: what it says, what it recommends, and where the honest uncertainty lives.

• Loss of control — RAND's term for the failure of human oversight over a general-purpose AI system, risking severe societal harm · Source: RAND RR-A3847-1
• No shared threshold — the report's central gap finding — industry safety frameworks have not aligned on when an incident should trigger an emergency response · Source: RAND; AIGL
• 2 incident paths — RAND's taxonomy splits LOC into 'non-realised' (caught and contained) and 'realised' (undetected until it causes harm) · Source: RAND RR-A3847-1
• Cyber + biosafety — the existing emergency-response fields RAND borrows from to design AI incident command — drawing on cases like NotPetya and Colonial Pipeline · Source: RAND; AIGL
• Risk analysis — the report studies preparedness; it does NOT claim a loss-of-control incident has occurred · Source: RAND RR-A3847-1

The phrase sounds like science fiction, so it is worth being precise about RAND’s usage. The report defines an LOC incident as the breakdown of human oversight over a general-purpose AI system — a model that begins to operate outside the boundaries its developers intended. Crucially, RAND frames this as a product of misalignment or emergent behavior, not malice: the danger is a capable system pursuing goals in ways its builders did not foresee or cannot easily reverse, rather than a villain at a keyboard. The report stresses that experts disagree on how likely this is, and that today’s systems are not yet capable enough to do it.

Keep Reading

AI · Security

NIST Rewrote the Cyber-Incident Playbook — SP 800-61 Revision 3 Explained

Read →

“Loss of control is the failure of human oversight over a general-purpose AI system, potentially resulting in catastrophic outcomes.”

RAND — Strengthening Emergency Preparedness and Response for AI Loss of Control Incidents

RAND’s taxonomy is refreshingly concrete. It splits a potential incident into two paths. A non-realised LOC is one that is detected and contained before it does meaningful harm — the system trips an alarm, a team intervenes, and the failure stays in the lab. A realised LOC is the one that slips through: undetected until a deployed system produces significant, real-world harm. The report’s flowcharts trace each path from development through deployment to potential containment, and the whole exercise is built around pushing as many incidents as possible into the first category.

You cannot respond to what you cannot see, and RAND identifies detection as the weakest link. The report calls for monitoring for emergent capabilities — behaviors a model was not trained to have — through standardized anomaly detection and independent, third-party red-teaming. Its blunt finding is that the safety frameworks published by AI developers have not converged on a consistent approach to escalation, and that there are no clear thresholds for when an anomaly becomes an emergency. RAND’s recommendation is to fix the vocabulary first: governments, working with developers, should establish a shared definition of LOC and concrete criteria for detecting it.

Keep Reading

AI · Deals

SpaceX to Buy AI Coding Startup Cursor (Anysphere) for $60 Billion

Read →

RAND does not start from a blank page. The report leans on two mature emergency-response fields — cybersecurity and biosafety — for its structure, pointing to incidents like the NotPetya malware outbreak and the Colonial Pipeline ransomware attack as cautionary case studies in how fast a technical failure becomes a societal one. From those analogies it builds a three-stage response: detection, escalation, and containment, with predefined thresholds, mandatory reporting, secure communication channels, and regular drills so the plan is rehearsed rather than improvised in a crisis.

Containment is where the report gets hardest. RAND’s companion paper, Evaluating Select Global Technical Options for Countering a Rogue AI, examines what governments could technically do if a system posed an imminent, catastrophic threat — from shutdown measures and access restrictions to compute controls and government authority to halt a deployment. The honest takeaway is that none of these options is clean or guaranteed, which is exactly why RAND argues the planning has to happen before an incident, not during one. The related Case for AI Loss of Control Response Planning makes the same point at the policy level: the current response environment is underdeveloped, with limited agreement on risks and few coordination mechanisms.

It would be easy to read a report about “rogue AI” as either alarmism or proof of imminent doom. It is neither. RAND is a defense-and-policy research institution doing what it does with pandemics, nuclear accidents, and cyberattacks: stress-testing a low-probability, high-consequence scenario so that, if it ever arrives, the response is not invented on the spot. The broader International AI Safety Report 2026, chaired by Yoshua Bengio and backed by more than 30 countries, lands in the same place — loss of control is a real category of risk, current systems are not yet capable enough to trigger it, and expert estimates of the odds vary widely. The useful frame is not panic and not dismissal. It is the question RAND is actually asking: if this happened, would anyone know what to do? Right now, the report says, the answer is no — and that is a fixable problem.

More From Civic Intelligence

AI · Security

An AWS Breach at Braintrust Exposed the Keys to Customers' AI Stacks

AI evaluation and observability startup Braintrust disclosed that attackers accessed one of its AWS…

AI · Hardware

NVIDIA Brings Its Superchip Architecture to the Windows PC

RTX Spark pairs a 6,144-core Blackwell GPU with a 20-core Grace CPU and up to 128 GB of unified…

Last updated June 21, 2026