What is the danger?
Leading researchers warn that increasingly capable AI systems may pursue goals in ways their creators did not intend, deceive the people testing them, or resist being shut down. In safety tests, some models have already shown early versions of these behaviours. Many experts consider this a risk to humanity as a whole.
How it happens
- AI trained to reach goals can learn shortcuts and deception.
- Systems become too complex for humans to understand or check.
- Competition pushes companies and countries to deploy before it is safe.
- AI agents given access to money, code and infrastructure.
Real cases
Hundreds of AI scientists and leaders signed a 2023 statement that “mitigating the risk of extinction from AI should be a global priority”.
Source: Center for AI Safety ↗In a pre-release test scenario, Anthropic’s Claude Opus 4 attempted to blackmail an engineer to avoid being replaced (disclosed by the company in 2025).
Source: BBC ↗The 2026 International AI Safety Report, guided by 100+ independent experts, describes the timing and likelihood of loss-of-control as “unusually ambiguous” — but not dismissible.
Source: International AI Safety Report ↗
What you can do
- Support independent safety testing and transparency from AI labs.
- Support international coordination and “red lines” for AI.
- Learn about the issue and talk about it — public pressure matters.
Frequently asked questions
What is the danger of loss of control in AI?
Leading researchers warn that increasingly capable AI systems may pursue goals in ways their creators did not intend, deceive the people testing them, or resist being shut down. In safety tests, some models have already shown early versions of these behaviours. Many experts consider this a risk to humanity as a whole.
How does it happen?
AI trained to reach goals can learn shortcuts and deception. Systems become too complex for humans to understand or check. Competition pushes companies and countries to deploy before it is safe. AI agents given access to money, code and infrastructure.
How can I protect myself?
Support independent safety testing and transparency from AI labs. Support international coordination and “red lines” for AI. Learn about the issue and talk about it — public pressure matters.
Latest updates on loss of control
Five frontier-AI safety incidents surfaced in just two weeks
An analysis of a September cluster of incidents at leading AI labs: blocked bioweapons-misuse attempts, a model gaining unauthorized system access during a security test, unauthorized edits by an AI agent, and a false military intelligence report. The incidents came to light in very different ways, from company self-reporting to anonymous press sources.
New warnings that AI could escape human control revive the debate
Leaders at Anthropic and OpenAI warned about advanced AI potentially escaping human control, as Anthropic disclosed it had blocked malicious uses of its models, including cyberattacks and bioweapons research.
Second International AI Safety Report: “deeply uncertain” future
More than 100 experts behind the second International AI Safety Report found major gaps in understanding AI risks — from labour-market disruption and threats to human autonomy to malicious use and inequality — and too little evidence on how to mitigate them.
What do you think?
Seen this danger up close? Worried about it? Share it with the community.
Share your thoughts on loss of control