COMPARISON

CanaryVaults vs Thinkst Canary

This page used to say the split was AI layer versus network layer. Thinkst announced Agent Provocateur on 4 September 2026, a decoy service that catches AI agents exploring a network, so that line is no longer true and has been removed. The difference that survives is about placement: their deception waits on infrastructure you control, for an agent that has arrived there. Ours travels with documents you have released, sits inline in front of your own models, and records what your own agents did. Many teams run both.

Start freeRead the docs

CAPABILITY BY CAPABILITY

What each platform is built for.

CapabilityCanaryVaultsThinkst Canary
Network / host decoysNot offered. No hardware or VM appliances.Core strength. Appliances and VMs that impersonate servers, network devices, and workstations.
Catching an agent that reaches your networkNot offered. Nothing we ship watches an internal network.Agent Provocateur, announced 4 September 2026: a decoy web service an exploring agent finds and interacts with, which then questions it about its task and access. Rolling out to customer Consoles and Canaries.
Decoy AI endpoints you publishDecoy chat endpoints you expose deliberately, with synthetic data, per-prompt intent classification, and visitor fingerprinting.Agent Provocateur is discovered by agents already inside a network rather than published as a public endpoint. Related goal, different placement.
Email decoy identitiesThree generated decoy addresses per protected email, seeded into fake credential dumps on paste sites. A contact records the mail server that delivered it and, when recognised, the platform targeted.Canarytokens cover related tripwire cases such as documents and URLs that phone home.
RAG / document trap factsFictional facts woven into documents; detection survives paraphrase and re-embedding, with per-copy attribution.Not a product focus; token-based document tripwires fire on open rather than on model ingestion.
Prompt-injection firewallThree-layer inline proxy (rules, semantic analyzer, sanitizer) adopted by changing one base URL.Not offered. Thinkst detects with decoys rather than filtering model input.
Audit trail for your own agentsSHA-256 hashed and sealed action log, each record linked to the one before, with public re-verification and PDF export.Not offered; alerting is the product surface, not agent action logging.
AlertingEmail and Telegram alerts with evidence attached; alert rules by intent class on honeypots.Mature multi-channel alerting with a strong reputation for low false positives.
Free tierYes. Usable limits on every surface except CanaryAudit, no credit card.Canarytokens.org is free; the Canary platform is a paid product (see their site for terms).

Capability summaries reflect each vendor's public documentation at the time of writing. We do not list competitor pricing. Check their site for it. If a row is out of date, tell us and we will fix it.

THE HONEST VERSION

Different problems, different tools.

Thinkst earned its reputation, and it is not standing still.

Thinkst Canary is widely respected for doing one thing extremely well: decoy devices that almost never false-positive. Agent Provocateur shows they are extending that into agent behaviour rather than treating AI as somebody else's problem. If your threat model is movement inside a corporate network, by a person or by an agent, they are the obvious first call.

The honest difference is placement, not which layer is 'AI'.

Agent Provocateur waits on infrastructure you own for an agent that has already reached it. Two of our surfaces are not on your infrastructure at all: trap facts leave with the document and fire when a model repeats them, and decoy identities are planted outside your perimeter and fire when someone contacts them. The other two sit inline in front of your models and around your own agents. Neither product covers the other's ground.

These tools stack.

Network canaries and AI-layer canaries answer different questions and do not conflict. Running Thinkst for infrastructure and CanaryVaults for the AI layer is a coherent setup, not a compromise.

CHOOSE THINKST CANARY IF

→Your priority is detecting lateral movement across an internal network.→You want to catch somebody else's agent once it reaches infrastructure you run.→You want physical or VM appliances that impersonate infrastructure.→You need a long enterprise track record for procurement.

CHOOSE CANARYVAULTS IF

✓You ship LLM features and need to know if documents leak into other models.✓You want a prompt-injection firewall in front of production models.✓You run autonomous agents and need a tamper-evident action log.✓You need detection that keeps working after a document has left your network.✓You want to start free today without a procurement cycle.

GO DEEPER

Field notes from the team.

A believable decoy teaches you more than a blocked request

Why the quality of the trap determines the quality of the signal.

Proof beats suspicion when private documents leak through AI

How planted trap facts turn a hard question into a binary signal.

Nobody detected the wiki incident. Someone reconstructed it.

Why a thing built to be found is the cheapest way to learn an agent is there.

Put canaries on your AI layer today.

Free tier, no credit card.

Start freeRead the docs