Transitive Taint Propagation for Shared Agent State: A Trust Primitive for the Verified Field

Autonomous AI agents are moving from coordination by message-passing between nodes to coordination through shared state—a common field of memory that many agents read from and write to. Every such system today verifies who writes (authentication) and manages when writes happen (concurrency), but none verifies whether a write is trustworthy before it becomes shared reality, and none traces the downstream consequences once a write is later found to be poisoned. We call this the unverified-writer gap. Because trust in a shared field is transitive—an honest agent may read a poisoned write and, in good faith, produce a derived write of its own—a single corrupted write can launder itself through honest agents and contaminate the field broadly before detection. This paper defines Transitive Taint Propagation (TTP): a primitive that (1) records, at write time, the provenance dependencies of each write as edges in a trust-dependency graph; (2) on detection of a poisoned write, computes its blast radius—the set of downstream writes that transitively depended on it—and (3) applies a containment operation that quarantines the tainted subgraph without requiring re-verification of the entire field. We specify the model, the propagation rule, and the containment operation; argue informally for the properties a correct implementation must satisfy (containment soundness, bounded over-containment, and tractability under a scale constraint); and state precisely what remains to be proven formally and demonstrated empirically. TTP is positioned as a normative extension to an existing execution-admission standard [2]; authentication of writers is treated as solved prior work and is not the contribution.