← All guides · Published November 14, 2026 · Reviewed November 14, 2026 · By the Ingenious AI team · 7 min read

Which AI Agents Cannot Be Blocked by Name?

The short answer: the three watch entries in AgentBlock's database: ChatGPT Atlas, Meta's Muse agent, and Grok retrieval. Their vendors publish no user-agent for them — Atlas browses as ordinary Chrome, Grok retrieval has been reported fetching like a regular browser — so there is no name to write a rule against. AgentBlock lists them transparently as watch entries, classifies their traffic UNKNOWN, and never guesses. What remains available is behavioural and contractual control, which this guide maps honestly.

Build my block kit — free · Test a policy in the simulator first

What is a watch entry?

an AgentBlock database entry for a real, reported AI agent whose vendor publishes no user-agent string. Watch entries — ChatGPT Atlas, Meta's Muse agent and Grok retrieval as of v2026-10-18 — cannot be emitted into block lists, are shown for transparency, and always classify as UNKNOWN.

No name, no rule: the mechanics

Every control in a name-based kit is a string match. The Worker compares the request's user-agent against the policy's substrings; a match draws the agent's verdict. An agent that sends a stock Chrome user-agent matches nothing — not because the kit failed, but because there is literally no string that distinguishes its requests from a human's browser. Any rule that 'blocked Atlas' by string would block Chrome, which is to say, your customers.

That is why the database gives these agents a formal status instead of silence. A watch entry records the agent's existence, its vendor, its class, its sources, and the reason it cannot be bound: no vendor-published user-agent as of the research date. The entry is the opposite of a gap — it is a surveyed boundary of what name-based control can reach.

The three, one by one

ChatGPT Atlas is OpenAI's browsing agent; the database notes it uses a standard Chrome user-agent, indistinguishable from a regular browser. Meta's Muse agent made the problem famous in September 2026, when Amazon blocked Muse through its Conditions of Use — a contract instrument — because, as reporting put it, robots.txt 'had no name to bind to'. Grok's retrieval is reported to fetch like an ordinary browser, with residential IP rotation and spoofed browser agents in the mix.

Notice the pattern in how each became known: not through vendor crawler documentation, but through reporting and directory research, which the entries cite. The database's research date and sources are printed per entry so an owner can see the evidence grade — watch is a statement about publishable names, not about whether the agent exists.

What UNKNOWN classification does

AgentBlock's classifier returns one of BLOCK with an agent ID, ALLOW with an agent ID, or UNKNOWN. Watch traffic lands in UNKNOWN by design — never guessed into an identity, never silently allowed-or-blocked by a fabricated rule. In evidence, UNKNOWN is a real category to watch: a rising UNKNOWN share with agent-like behaviour patterns is the signal that the unnamed population is growing on your site, which is information a policy review can act on even without names.

The classify self-test in the codebase enforces this discipline mechanically: stock browser strings, an empty agent, and even directory-reported tokens like GrokBot all classify UNKNOWN. The boundary is tested, not aspirational.

The controls that remain

Behavioural controls do not need names. Rate behaviour applies at the edge regardless of identity; path fences protect your private subtrees from everything not authenticated, named or not; canary paths detect structural mapping by unnamed visitors; and authentication itself remains the real boundary for anything genuinely private. Contractual controls operate at the other layer entirely — terms of use that prohibit automated access create the legal posture Amazon used, which matters to institutions even when it stops no packets.

What owners should not do is buy certainty theatre: tools that promise to block these agents by fingerprint heuristics without disclosing error rates. A false positive against a watch agent is a blocked customer. AgentBlock's position — printed in its database design — is that unknowns are reported as UNKNOWN, and controls are sold only where they can be honest.

Will watch entries ever become blockable?

Possibly. Vendors do publish agent documentation after launches, and the database is versioned for exactly this: an entry that gains a vendor-published user-agent in a later version moves from watch to verified and becomes a policy decision like any other. The stale check is the mechanism that would surface the change to existing kit owners — one more reason to regenerate on flags rather than treating a kit as a one-time installation.

Until then, the mature posture is the one this guide models: know the three names, know why they are unbound, keep behavioural fences in good repair, and let your policy documents say 'these agents are outside name-based control' in plain words. Precision about a limit is worth more than a confident bluff past it.

Five actions to take today

  1. Learn the three watch names. ChatGPT Atlas, Meta's Muse agent and Grok retrieval — real agents, no published user-agents, as of database v2026-10-18.
  2. Never accept a guessed rule. A block string for an unnamed agent is either fiction or a rule that hits ordinary browsers; AgentBlock emits neither.
  3. Fence privately by path and login. Behavioural and authentication boundaries are the controls that bind unnamed traffic.
  4. Watch the UNKNOWN share. Track unidentified agent-like traffic in reviews; growth there is your cue to strengthen behavioural layers.
  5. Regenerate on database updates. If a watch entry ever gains a published user-agent, a fresh kit converts it into a normal decision.

Worked example: a marketplace answers the unanswerable question

Harvest Lane Marketplace, a fictional marketplace for farm producers, was asked by a large seller — in writing — whether the platform 'blocks Grok and Muse from scraping our listings'. The honest answer required the watch-entry facts: neither agent publishes a user-agent, neither can be blocked by name, and any vendor claiming otherwise is guessing.

What Harvest Lane could document was its behavioural posture: listing APIs behind authentication, catalogue paths rate-shaped at the edge, a canary path in the kit, and terms of use prohibiting automated extraction. Its evidence pack for the quarter showed 96% of listing-page traffic classified to named agents or ordinary browsers, with the UNKNOWN remainder counted and reported rather than hidden.

The seller renewed. The marketplace's account manager later said the deciding factor was not a promise but the precision — a platform that knew exactly where its controls ended was trusted with the rest.

Frequently asked questions

Why doesn't AgentBlock just block the names people report for these agents?

Because reported names are not vendor-published names. Directory tokens like GrokBot are explicitly excluded from block lists — a rule bound to an unverified token is a guess, and guesses in production policy either do nothing or block the wrong traffic while reporting success.

Can robots.txt block Atlas, Muse or Grok?

Robots rules need a name to bind to as well. That is precisely why Amazon's Muse response ran through its Conditions of Use instead. Terms and contracts remain available instruments; they create obligations and evidence, not packet-level refusals.

If they look like browsers, is my analytics lying?

Your analytics is reporting user-agents faithfully; it is the agents that are indistinguishable at that layer. Treat UNKNOWN and browser-class traffic with agent-like patterns as a category to monitor behaviourally — volume shape, path patterns, canary hits — rather than by name.

Will these agents stay unblockable forever?

Unknown — and the database is built for the change. Entries are versioned and dated; if a vendor publishes a user-agent, the entry can graduate to verified in a later database version, and your stale check will flag that your kit predates the change.

What is the single most useful control for watch-class traffic?

Path fencing plus authentication on anything non-public, because it does not depend on identity at all. After that, behavioural rate shaping and the canary path give you detection and evidence for the public surface.

Build my block kit — free · Browse the agent database

Related guides

HONEST SCOPE: this policy blocks KNOWN agent user-agents and behaviors listed below. Sophisticated spoofing can bypass any user-agent block (e.g. an agent sending a stock browser UA, or rotating residential IPs). This is a control, not a guarantee.
← Back