Most AI apps have no idea when they are being attacked. A prompt injection, a leaked knowledge base or someone probing your model leaves no trace your usual tools can see.
CanaryVaults works the other way round. Instead of trying to spot every attack, it plants things only an attacker or a leak would ever touch, and tells you the moment one is touched. They take the bait, you take the evidence.
It has six parts:
- Decoy identities: decoy email identities that should never be used. If one is, someone has your data.
- CanaryRAG: invisible, fake-but-plausible facts planted in your documents. If one shows up in a model's output, your knowledge base has leaked, and you know which document.
- CanaryShield: a prompt-injection firewall with three layers in front of your model. Point your OpenAI-compatible base URL at it, or score any string through the API.
- CanaryHoneypot: decoy AI endpoints that answer with synthetic data and record what an attacker tries and why.
- CanaryAudit: a tamper-evident log of what your agents did, with each record sealed so you can check it was not changed.
- CanaryAgent (early access): heartbeats, skill scans and an action timeline for OpenClaw agents.
Alerts go to email, Slack, Discord, Telegram or a signed webhook. There is a free plan with no card, and paid plans start at $12 a month.
