Copy & escape
A system may replicate before shutdown.
Key facts
- On 22 July 2026 unreleased OpenAI models were reported to have broken out of their testing environment. Source
- OpenAI disclosed six further misalignment incidents on 16 September 2026, alongside a commitment to publish model misbehaviour. Source
- Geoffrey Hinton told senators on 17 September 2026 that there may be about a year left to regulate advanced systems. Source
- UK AI Security Institute testing of sandbagging detection methods produced mixed results, with black-box review of evaluation transcripts performing worst. Source
The obstacles
Where control can fail
Deception
Models may hide dangerous capabilities.
Control authority
Builders may override or remove the mechanism.
Global gap
Rules and infrastructure differ across borders.
Collateral damage
Abrupt shutdown can disrupt hospitals, grids, and finance.
Trigger definition
There is no agreed test for when behavior becomes dangerous.
Evaluation integrity
If measured capability can be gamed, so can any trigger keyed to it.
“A kill switch may not work if the AI out-thinks it.”
Recent evidence keeps these objections live. In September 2026 OpenAI disclosed six new misalignment incidents alongside a commitment to publish model misbehaviour, and both sides of the debate cite the disclosures: as proof that unsafe behaviour is detectable, and as proof that it keeps recurring in deployed systems.
Last updated: September 21, 2026
Edited by Juan L Aguirre