In plain English
This page explains why testing AI parts one by one is necessary but incomplete. Safe-looking parts can still produce unsafe behavior when combined.
- Why this matters: AI risk can come from the whole arrangement, not one obvious model.
- What to look for: data, memory, routes, adapters, tools, evaluators, updates, and rollback paths.
- Technical version below: the expert terminology remains available and is linked through the glossary.
Merge-State Backdoors and Plausibility Camouflage
The report corpus repeatedly highlights model mergingCombining model weights or adapter deltas into one artifact. Open glossary definition and adapter composition as a supply-chain blind spot. A dangerous behavior may be distributed across components that look useful and benign when inspected separately.
Plausibility camouflage
A component can perform well on an ordinary task, preserve apparent utility, and pass routine checks. That usefulness is the camouflage. The unsafe behavior is not necessarily in the component alone; it may appear only when another component shifts the representation, refusal boundary, or routing condition.
Why merge state matters
A merge is not just a file operation. It creates a new model state with its own behavior. The merge coefficients, load order, base family, quantization, and prompt policy must be recorded as part of the evaluated artifact.
Safe framing
This page does not provide backdoor construction guidance. It explains why defenders cannot rely on component reputation alone.