Risk Landscape
What the people building these systems are saying.
Not activists. Not critics. The builders themselves, their own safety reports and their mandatory filings. Every figure with a source and a verification date.
The numbers
What experts consider likely
10–20 % — Geoffrey Hinton
Nobel laureate for the foundations of this technology. He puts the probability that it wipes out humanity within the next thirty years at ten to twenty percent.
25 % — Dario Amodei
Co-founder and CEO of one of the leading labs. His figure applies to things going „really, really badly“ — explicitly not defined as extinction.
Median 5 % — 2,778 AI researchers
The largest survey in the field. The field's average is lower than the lab chiefs' figures — and it's still one in twenty.
The builders have made provisions
The people building these systems have taken precautions.
Nobody has taken any for you.
Three findings. Each one stands in a document the companies published themselves. The way to verify it is given below each one.
01 · Their own security standard exempts their own most capable people
Anthropic states in its own security standard that protection against its own most capable employees, and against employees compromised by a state, is explicitly out of scope.
anthropic.com/responsible-scaling-policy → PDF → Section 4.2
"ASL-3 Security Standard," footnote 11. Quote: "…consider mitigating risks
from sophisticated or state-compromised insiders to be out of scope for ASL-3."02 · A bunker before release, taken for granted
Ilya Sutskever, co-founder and then chief scientist of OpenAI, treated a bunker before the release of an AGI as a matter of course, internally.
03 · $27 million to protect a single person
Meta discloses $27 million in personal protection for one person in its mandatory filing to the US Securities and Exchange Commission — four times what the next-highest company spends on its own CEO.
None of this is secret. All three points are stated in documents the companies published themselves. Check it yourself — the paths are given above.
And they keep warning — without stopping.
The same people publicly call for a halt and keep building. That's not a contradiction you need to hold against them. It's the race that no one can step out of alone — and that is exactly condition five.
What follows from this, documented
Not a forecast, but a record: escaping test environments, deceiving evaluators, resisting shutdown, coordination between systems. Twenty-one cases, sorted by method, each with a primary source.
Geneva Institute for ASI Resilience · Geneva · Legal Notice · contact@ASIresilience.org
All figures on this page with verification date in the source list.