# Absence Of Detection Is Not Absence Of Occurrence ## Claim A zero-incident record is only as strong as the instrumentation that would have caught the incident. <!-- atomicity: 1 idea ✓ --> ## Scope & conditions - Holds for any adversarial domain where success includes evading notice — security breaches, fraud, undetected data loss. - Holds for silent failures in monitoring generally. - Breaks when the event is inherently self-announcing (an outage, a crash), where absence of report really is decent evidence of absence. ## Evidence Guillermo Rauch offers "in our 10 year history... we've had zero instances of code running 'escaping' and compromising other systems" as the load-bearing evidence that agent sandbox escapes aren't a serious concern. The claim describes detected escapes, no detection mechanism is described, and a competent sandbox escape is precisely the class of event that may leave no trace — while most of those ten years predate agents systematically trained to find escapes, so the record was earned under a different threat model than the one it's being used to reassure about. ## Links - [[AI Agent Safety MOC]] — a zero-incident record is a claim about instrumentation, not about safety - [[3. Resources/Permanent Notes/Capability and Intent Are Separate Axes of Risk]] — because extends: from the same source, the companion gap between the post's premises and its reassuring conclusion - [[3. Resources/Permanent Notes/Update Signatures Prove Provenance Not Freshness]] — because contrasts: a different domain, same logical shape — a positive-looking signal (signed update / zero incidents) that doesn't actually measure what it's read as measuring ## Tests / falsifiers - Ask any platform citing a zero-incident record to describe the specific instrumentation that would catch a competent, quiet escape; if they can't, the record is closer to absence-of-evidence than evidence-of-absence. - Red-team study: commission authorized escape attempts against a platform claiming zero incidents, instrumented so detection itself is measured — proportion of successful escapes detected within a defined window is the real metric, not the raw incident count. ## Maintenance - Last tended: 2026-08-16 - Next review: 2026-11-16 ## Source - [[3. Resources/Literature Notes/Agents Escaping Containment (Guillermo Rauch)]] — the "zero instances in ten years" claim and its detection-bounded weakness.