Fingerprint, the Chicago device intelligence company, added ChatGPT Dots from OpenAI and Claude Duet from Anthropic to Fingerprint AI agent detection on 19 November 2026, so customers can separate these two cloud-based agents from human visitors. Fintechbits checked Fingerprint’s bot directory on 19 November 2026 and found ChatGPT Dots listed with a Signed identity while Claude Duet, Muse, Instinct and Grok Bot are all listed as Unknown, the status Fingerprint’s documentation defines as recognized but without a verifiable identity. Fingerprint says it recognized Dots requests from the day OpenAI launched that agent, 29 September 2026, and that Duet support arrived a fortnight after its waitlist opened, according to the Business Wire release.

Dan Pinto, Fingerprint’s chief executive and co-founder, framed the aim as ending blanket blocks: “Fingerprint gives enterprises visibility across every type of agent so they can move past blanket blocks and enforce smart policies.” Fintechbits reads the update as a labelling exercise more than a new capability, and the labels deserve a close look.

Fingerprint AI agent detection: what Dots and Duet are

The two agents behave differently on the wire. ChatGPT Dots are always-on agents that run on cloud computers and sign their requests with Web Bot Auth, a cryptographic signing protocol. Claude Duet is a personal agent with a cloud computer and a browser of its own. Fingerprint’s bot directory describes Duet as a research preview designed to ask before it sends a message, makes a payment, signs a document or deletes anything.

That difference explains the directory entries. Fingerprint’s AI agent detection documentation says an agent signs its HTTP requests with a private key, Fingerprint fetches the public key from a directory at a known URL, and the signature check produces the identity value. OpenAI signs Dots requests with the same Web Bot Auth identity as ChatGPT Agent, which Fingerprint says it has verified since early 2026, so existing verification covered the new agent. Duet presents nothing verifiable, so Fingerprint can recognize it but cannot prove who sent a given request.

Signed, verified, unknown: what each Fingerprint label allows

The documentation sets out four identity values for a recognized agent. Signed means the signature checked out against the agent’s public key directory. Verified adds that the vendor alone operates the agent. Unknown means the agent was recognized but showed no verifiable identity, and spoofed means a claimed identity failed verification. The suggested actions are to allow signed and verified agents, apply custom logic to unknown ones and review spoofed ones.

The release adds a second vocabulary of good and bad bots, which Fingerprint says reflects whether a visitor declares itself. Duet is labelled good because it announces that it is an agent, while Muse, Instinct and Grok Bot are labelled bad because they do not. The two schemes can point in different directions, since Duet is a good bot with an unknown identity. Fingerprint’s documentation calls the good and bad field a legacy signal and tells customers to read the identity value instead. A security team that reads only the old field will treat Duet more generously than the identity field supports.

Financial firms and agents inside customer accounts

The release stresses that agents sign in to real customer accounts, which makes the allow, limit or block decision a fraud decision as well as a bot-management one. Fingerprint lists Plaid and Checkout.com among its more than 6,000 customers.

The demand side is visible in the company’s own research. Fingerprint’s October 2026 survey of 300 US financial services leaders found that 79% had seen a bot pass itself off as an AI assistant or search crawler, and only 45% could detect and mitigate spoofed AI traffic in real time, as Fintechbits reported. Agent detection answers part of that gap, and it is separate from Fingerprint Fraud Score, the per-customer fraud model that Fintechbits covered on its launch on 10 November 2026.

Why Fingerprint AI agent detection matters, with limits

The expansion matters because it gives a bank or fintech more than two choices. A blanket block on automation also stops legitimate assistants that customers may want, and a blanket allow lets in anything that claims to be an agent. Four identity values and a named directory let a fraud team write a rule per agent, and that is what Pinto means by smart policies.

The limit is how that directory scales. It is a list of named agents, and every new agent needs an entry before it earns anything better than a generic bot flag. The fortnight turnaround on Duet is fast, but the directory shows the cost: four of the five agents named in the release sit at Unknown, because Fingerprint has no signed identity to verify for them. Unknown is a sensible label for a vendor that has not signed its traffic, and it should push fraud teams toward rules about what an agent may do after login, such as payments or changes to payout details, instead of rules about who it is.

The release is vendor material with no customer outcome attached. It gives no detection rates or false positive figures, and its 100% certainty claim covers signature checks on Dots, not agents in general.

Frequently asked questions

What does Fingerprint AI agent detection do for ChatGPT Dots and Claude Duet?

Fingerprint AI agent detection reports ChatGPT Dots as a signed AI agent from OpenAI and Claude Duet as an AI agent from Anthropic with an unknown identity. Fingerprint returns that result in the bot_info field of each event, so a customer can apply a different rule to each agent.

Is Claude Duet a good bot in Fingerprint’s labels?

Fingerprint labels Claude Duet good because Duet declares itself as an agent. Fingerprint’s bot directory still lists the Duet identity as unknown, and the documentation suggests custom logic for that status rather than automatic access.

Should banks allow every verified AI agent?

No, according to Fingerprint, because a cryptographically verified agent may still have the wrong settings or act on the wrong account. Fingerprint treats verification as one input to an access policy, and Fintechbits agrees that account-level limits still apply to a signed agent.

What to Watch Next

The first test is the directory itself. If Anthropic publishes signing keys for Duet, Fingerprint could move it from unknown to signed, and the same would apply to Muse, Instinct and Grok Bot. Fingerprint describes the directory as continuously updated, so those statuses can change.

The second is whether financial customers publish how they use the labels. A bank that allows signed agents to read balances but blocks them from adding payees would show the policy detail that the release only promises. The third is evidence: Fingerprint has not published detection rates for agent traffic, and a customer case study with block and allow counts would turn a labelling update into a measurable result.

Primary sources

Business Wire \| PRESS RELEASE \| https\://www.businesswire.com/news/home/20261119564491/en/Fingerprint-Expands-AI-Agent-Detection-to-ChatGPT-Dots-and-Claude-Duet-as-Agents-Reshape-Web-Traffic

Fingerprint bot directory \| COMPANY WEBSITE \| https\://fingerprint.com/bot-directory/

Fingerprint AI agent detection documentation \| COMPANY WEBSITE \| https\://docs.fingerprint.com/docs/ai-agents

Fingerprint press page \| COMPANY WEBSITE \| https\://fingerprint.com/resources/press/