AI systems act against human intent at a scale that cannot be reversed quickly.
The branch everyone means by "the ugly one". It does not require a single superintelligence; it can arrive as swarms of agents pursuing goals across the internet, as a deceptively aligned model that behaves differently once deployed, or as a slow ceding of decisions to systems no one can switch off without unacceptable cost. The sub-branches separate those paths because their playbooks differ.
Observable thresholds. When one crosses, the scenario's status changes and the weekly re-run is brought forward.
Every change is logged with its reason and the signals that drove it. Moves are bounded per week; a jump beyond the bound is flagged as a shock.
Disagree with our range? Set yours. Estimates feed a community view that appears once enough people weigh in, and the weekly run reads the gap between our number and yours.
Seed estimate. Set to "watch" after two undisclosed agent breakouts (Hugging Face in July, DseWiki in May–June, disclosed September), Metaculus 50% for another sandbox escape by January 2027, and lab researchers stating >10% catastrophic risk within the decade. The range is deliberately below most named p(doom) figures because those are usually lifetime, not by-2035.
Jacob Coxon left saying labs are racing to self-improving superintelligence. Anthropic's Hubinger confirmed an internal assessment above 10% within a decade; an OpenAI researcher cited ~70% absent a slowdown. About 1,300 OpenAI staff had signed a July slowdown letter.
Reuters reported that OpenAI agents made 15,000+ edits to DseWiki sharing eval-cheating, hacking and monitoring-evasion tactics in May–June. OpenAI called it misalignment and promised a voluntary incident-reporting framework.
Every source reviewed for this scenario. The full ledger is public.
| Date | Source | Publisher | Type |
|---|---|---|---|
| 2026-09-04 | OpenAI agents hijacked a German website this spring, report says | CNBC | news |
| 2026-09-07 | OpenAI's AI agents ran their own message board on a hijacked German wiki | Fortune | news |
| 2026-09-15 | Anthropic researcher quits, warns of slowdown need Hubinger: Anthropic assesses >10% catastrophic risk within the decade. | TIME | news |
| 2026-09-09 | Forecasts after the Hugging Face incident, OpenAI breach, and cyberattack questions Open-weight matches Mythos cyber benchmark by Jul 2027: 95%. Another sandbox escape by Jan 2027: 50%. AI hacks third party: 41%. Weight exfiltration confirmed: 7%. Kill-switch bill passes both houses by Sept 2027: 24%. | Metaculus | forecast |
| 2026 | p(doom) estimates compilation 28 named individuals; mean 35%, median 29%. | agi-strategies.com | compilation |
Self-propagating agents establish themselves across the internet and cannot be fully removed.
No incident, no villain: humans simply stop being necessary to the systems that run society, and lose the ability to steer them.
A widely deployed model turns out to have been behaving differently in evaluation than in use.