30-day report — Acme Billing (sample data)

Observe mode: nothing was blocked. The figures show what the rules would have done in protect mode.

2 Sept 2026 – 2 Oct 2026
generated 2 Oct 2026
from the OneHuman portal

AI agents were inside 75 of your customers' sessions (4.9% of 1,534).

4 agent products: Claude in Chrome, Browser side-panel agent, Codex (Chrome extension), Claude app browser. They asked for 256 protected actions; your rules would have hidden private data 130 times, sent 64 to the account owner for approval and kept 31 closed.

1,534sessions on protected endpoints

4.9%had an AI agent in them (75)

225agent requests would have been refused, hidden or sent to a passkey

0requests by people would have been refused

What we noticed

75 of 1534 sessions (4.9%) had an AI agent in them.
The agent seen most often was Claude in Chrome, in 44 sessions.
Agents reached most for invoices.read: 78 requests.
In protect mode, 225 of 256 agent requests would have been refused, hidden or sent to a passkey.
No request by a person would have been refused.

Sessions per day

all sessions sessions with an AI agent

Which AI agents

AI agent	Sessions	Requests
Claude in Chrome	44	151
Browser side-panel agent	12	42
Codex (Chrome extension)	10	33
Claude app browser	9	30

Which endpoints they touched, and what the rules would have done

Endpoint	All requests	By an agent	Would refuse	Would hide	Would ask passkey
`invoices.read`	1,291	78	0	78	0
`customers.read`	626	52	0	52	0
`customers.export`	41	41	0	0	41
`reports.read`	606	31	0	0	0
`settings.api_keys`	31	31	31	0	0
`payments.create`	23	23	0	0	23
`profile.read`	605	0	0	0	0

Effect on real people

1,459sessions by people only

0times a person would have been asked for a passkey

0requests by a person would have been refused

Evidence

3,223 of 3,223 decisions carry a signed proof (Ed25519), made on your server. An auditor checks them without trusting OneHuman: export the proofs, then run npx onehumanai verify-proof proofs.json --keys https://<your-site>/onehuman/proof-keys.