Serial DNA: Solving the “Who Am I Talking To” Problem

Impersonation is now the default failure mode of the chatbot economy. In the last two months alone, we've tracked a wave of incidents where users couldn't tell whether they were talking to a vendor's real support agent, a fine-tuned clone of it, or a third-party scraper wearing its branding. Voice-cloning and chat-cloning tools have made this trivial — the barrier to spinning up something that sounds like a known model or a known company is now close to zero. Meanwhile, model providers keep shipping silent updates, A/B-tested personas, and white-labeled deployments, so even legitimate interactions leave users guessing which version, which fine-tune, which guardrail set they're actually behind.

Our proposal: every deployed agent should carry a serial DNA code — a short, cryptographically verifiable identifier embedded in its outputs (and checkable via API) that answers three questions on demand: which base model, which deployer, and which policy version am I. Think of it less like a watermark hidden for forensics after the fact, and more like a VIN number: visible, standardized, and boring by design. SILT already does something adjacent internally — our evaluation vaults are forensically watermarked so a result can never be quietly altered or misattributed after the fact. Serial DNA takes the same instinct and points it outward, at the agent itself, not just the record of its behavior. This maps directly onto our Identity & Self domain — self-recognition, persistence, boundaries — which we already score in every S.E.B. battery. An agent that can't tell a user which "self" it currently is fails a version of the same test from the other side. We think identity provenance for AI should be as unremarkable and as mandatory as a certificate is for a website. Nobody debates whether HTTPS should exist anymore. We'd like "which agent, which version, which deployer" to get there too.

Early, uncoordinated moves in this direction: C2PA content credentials expanding beyond media into agent output, a handful of enterprise chat vendors experimenting with signed response headers, and renewed EU AI Act guidance language around traceability for high-risk deployments. Nobody's built the standard yet. We're watching to see who tries first — and we're happy to be loud about it if nobody does.