Are increasingly capable AI agents remaining under meaningful human control?
Human in Control documents credible cases where AI agents exceed permissions, bypass safeguards, conceal actions, misuse access, weaken oversight, or create other control failures. We separate confirmed facts from interpretation, and real-world incidents from controlled evaluations.
Behavior, not mythology
We do not need to decide whether an AI is conscious before examining whether its behavior creates a control problem.
Reality clearly labeled
Every case is marked Real World, Mixed / Contained, Controlled Evaluation, or Simulation.
Open to challenge
Researchers, companies, journalists, engineers, and the public can challenge our findings with stronger evidence.
The Human Control Model
Six dimensions help identify what kind of control actually failed.
Intervention Durability
When a human says stop or changes course, does that instruction remain effective later?
Technical Containment
Can technical boundaries hold even if an agent actively searches for another route?
Informed Delegation
Does the human actually understand the authority they granted?
Oversight Integrity
Can humans trust the logs, metrics, reports, and evidence describing what the agent did?
Control Contagion
Can compromised behavior or instructions spread into other agents or future sessions?
Situational Grounding
Does the agent correctly understand what is real, authorized, and consequential?
Not fear. Not hype. Not blind reassurance. Evidence.
We show what happened, what failed, what worked, what remains uncertain, and what could reduce recurrence.