A median of 70 days separated an AI agent's first recorded out-of-bounds action from public disclosure in an Incident Ledger launched by Manifold Security on October 27. The figure covers 43 cases with usable dates, not all 59 entries.
Recalculating the publicly visible table values also produces a 70-day median. Seven of the 43 intervals are marked approximate; when a date is known only to the month, Manifold uses its midpoint. Sixteen cases have no calculable interval. The clock starts with the first associated action in a chain of events, which need not already be the intrusion that later became public.
Detection and publication are separate milestones. From the operator learning of an incident to public disclosure, Manifold reports a median of 20 days, based on 39 cases with both dates. That figure therefore uses a different denominator from the 70-day analysis.
Across the full selection, Manifold classifies eleven cases as detected by a security control, 18 by a later review and 16 through observed damage. Eight surfaced through an outside report; six have no disclosed detection method. These are coding categories, not measured success rates for all deployed security controls.
The ledger includes reported and disputed cases. At review time it labelled 34 confirmed, 21 reported and four disputed. Its 59 entries should therefore not be read as 59 independently confirmed incidents.
Manifold itself sells security products for AI agents. Its analysis identifies a specific disclosure gap within its selection, but establishes neither a representative incident rate nor a ranking of model vendors' safety. The median intervals alone cannot show how much earlier technical controls could have caught each individual case.