Disclosure lag
Could a lab hide AGI? One piece of evidence is how long real AI incidents stayed out of public view. Each bar runs from when an incident happened to when it was made public.
Days from incident to public disclosure
What counts
An incident is listed if an AI agent took an unsanctioned action outside its sanctioned environment that touched a real third-party or public system, it was made public by anyone, and it's backed by a primary source or two credible outlets. There's one entry per first public disclosure. Where a source gives only a month or season, we record the window and measure the lag from its midpoint; the range on the median shows the ends of those windows.
What this can't show
Only incidents that became public appear here. Anything undetected, or detected and still hidden, is missing by definition, and the tracker only covers incidents from 2026 on, so it can't yet show a secret kept for years. Read these lags as a floor on how long things stay hidden, not a typical or a maximum value. The lag also runs from when an incident happened, so it mixes time before anyone noticed with time when a lab knew and said nothing; where a source says when the lab detected it, the chart marks that date.
Why it matters
This is the closest thing we have to an empirical base rate for hypotheses A and C: how long real AI behaviour stayed out of public view, and who brought it to light. In 2026 the incidents that did come out took from days to months to surface, and several were made public by someone other than the lab: victims, outside researchers, evaluators, governments or reporters. That shows sub-AGI agents can act out of public view for months, and that labs don't reliably disclose first. It can't show how long a careful system, or a lab set on keeping a secret, could stay hidden. New incidents are added in each weekly wrap-up.
Separate, lab-reported context: Anthropic says that in August 2026 about 30,000 research and engineering agents ran at once on its most-used internal platform, every action passed through an automated monitor before it ran, and about 1 in 47,000 actions was blocked (Anthropic). That is self-reported and not independently audited, and it measures what one lab's monitor catches, not what reaches the public.