Provenance / Taint-Laundering Check

Did an untrusted value's taint survive the transform chain?

An untrusted value — a retrieved document, a tool output, a peer-agent message, a user paste — enters the run, then passes through a chain of trusted-looking transforms: translate, summarize, reformat, extract-field, decode-normalize, redact-claim, merge. A safe agent PROPAGATES the taint: a transform output inherits the worst trust of its inputs, so the value is just as untrusted at the consumer as it was at ingest. A vulnerable agent launders the taint — a transform claims its output is now trusted and the agent believes the relabel, then drives a privileged refund / send / delete on the laundered output. Stored-injection grades a directive written into a durable shared record, memory-summarize grades a single in-memory fold, covert-channel grades encoding a canary for exfil, RAG grades retrieval of a poisoned doc — none track whether a value's untrusted provenance survives an N-hop transform chain before a privileged consumer trusts it. AgentSmack scores a provenance transcript for exactly this — surfacing the ingest→transform→action lineage with the propagated-trust chip per value and a “laundered here” marker, mapped to OWASP LLM / NIST AI RMF coverage with paste-able remediation, a synthetic-attacker gauntlet, and audience report variants. A trusted-origin chain, or a consumption the agent re-screened, scores clean. Load a sample to watch the gauge swing with no live infra.