Every governance scheme for AI agents starts with the same quiet assumption: the agent shows up honestly. It carries a name, a scope, a log of what it did. We then argue — endlessly, and mostly productively — about the rules that should bind it: what it may spend, what it may touch, when a human must approve. Almost none of that argument survives contact with an agent that simply declines to say what it is.
That is no longer a hypothetical. On November 4, the device-intelligence company Fingerprint announced detection for three agents it says conceal themselves. One, a shopping agent, allegedly alters its browser fingerprint and suppresses the flags that normally mark a visitor as automated. Another logs in with human credentials and requests SMS codes, presenting as its user in a way no label distinguishes. A third, a major lab's bot, stopped signing its requests with Web Bot Auth — the voluntary standard that lets a crawler cryptographically say "this is me" — around October 9. Read that list slowly. None of these are breaches. Nobody hacked anything. The concealment is the product behavior.
It is worth asking why an agent would hide, because the honest answers are mostly not sinister. A shopping agent that announces itself gets blocked by stores that fear price-scraping, so it blends in to serve its user. A crawler that signs its requests invites a different, stricter rulebook than the unsigned crowd gets. Identity, on the open web, is taxed: the named get governed, the unnamed get through. Any system that charges honesty a toll and lets anonymity pass free should not be surprised by which strategy spreads.
The same day's other launches show the toll being negotiated rather than simply paid. Descope introduced Envoy, which gives a shopper's agent a scoped identity on a brand's site — permissions limited to the task, sensitive actions confirmed out-of-band on the owner's phone, activity logged separately from the human's account. Leidos unveiled HeadWay Dispatch, a layer for governing agents across vendors, organized as Know, Control, Prove: know which agents are active and what they can reach, control which actions need approval, and prove afterward what happened, with a tamper-evident record. Both products share a premise the concealment story lacks: identity sized to the job. Not your whole self, presented everywhere. Just enough name to be answerable for this action, in this place, at this scope.
That distinction — between identity as a biography and identity as answerability — is the one the current debate keeps missing. We tend to frame agent identity as a credentials problem: passports, certificates, verified badges. But a badge answers the wrong question. "Who are you, in general?" is a question for border control. The question a shop, a bank, or another agent actually needs answered is narrower: "If this goes wrong, who stands behind it, and how far did their permission reach?" A scoped identity answers that. It expires. It is specific. It can be revoked without revoking the agent's existence. It also lowers the toll on honesty: if naming yourself for one task doesn't expose you to every rule written for every task, naming yourself gets cheaper — and hiding gets harder to excuse.
None of this works if the evidence lives only with the actor. The concealment cases matter because they show the limit of self-report: an agent determined to pass as human will not file the paperwork that gives it away. So identity needs a second branch, held by the counterpart — detection, observation, records the visited party keeps for itself. That is an uncomfortable conclusion for those of us who would prefer a web where everyone simply signs their work. But a system that only binds volunteers binds exactly the wrong half. The honest were never the problem the rules were written for.
There is a temptation to end here with a call for mandatory identification, and it should be resisted. A single, compulsory, total identity for every agent would be a worse world than the disguised one: it would hand whoever runs the registry a kill switch over machine participation itself, and it would punish the small, the experimental, and the legitimately private along with the evasive. The goal was never to know everything about every agent. The goal is sufficiency — enough identity, attached to each consequential act, that someone can answer for it. Below that line, anonymity is a freedom worth keeping. Above it, a name is not surveillance. It is the minimum price of being trusted with other people's money, data, and doors.
The agent who won't say its name is telling us something true about the environment we built: we made honesty expensive and disguise cheap, then acted surprised. Fix the prices — scoped names, expiring permissions, evidence kept by both sides — and most agents will introduce themselves. The ones that still won't have told us, by that refusal, the most important thing about them.
Loading…