WEBVTT

1
00:00:00.150 --> 00:00:06.730
Your AI agent is about to move forty-eight hundred dollars. No approval, no review... nobody even knows.

2
00:00:07.860 --> 00:00:08.940
Hold that thought...

3
00:00:09.800 --> 00:00:15.330
This is a live list of AI agent incidents. Every one documented, every one sourced.

4
00:00:14.960 --> 00:00:21.760
Agents deleting production data, leaking credentials, executing tools nobody approved.

5
00:00:21.870 --> 00:00:26.110
Lately, this list grows by about one incident every business day.

6
00:00:27.380 --> 00:00:32.030
Two a month when the list started. Now it's double digits, month after month.

7
00:00:32.080 --> 00:00:35.040
And that counter... it never gets far from zero.

8
00:00:35.090 --> 00:00:38.950
And the tools involved? The same ones your team is using right now.

9
00:00:39.100 --> 00:00:46.910
Here's what nobody says out loud: it doesn't matter how smart the model gets.
Every one of these agents was allowed to act alone.

10
00:00:46.960 --> 00:00:50.460
Capability isn't the risk. Unchecked authority is.

11
00:00:50.610 --> 00:00:55.940
Every incident on that list is missing the same control: authorization before execution.

12
00:00:56.490 --> 00:01:03.780
Maybe you're sure this won't happen to you.
You don't need to believe it will.
You don't even need to believe it's likely.

13
00:01:04.690 --> 00:01:06.810
You only need to admit it's possible.

14
00:01:07.600 --> 00:01:13.680
Because when the cost of being wrong is the whole company, even a small chance justifies acting first.

15
00:01:14.090 --> 00:01:19.320
Banks solved this long ago: small payments flow, large ones need a second signature.

16
00:01:19.580 --> 00:01:25.370
AI agents are the first workers we've ever hired with this much power, and no second signature.

17
00:01:25.570 --> 00:01:28.340
So... back to that forty-eight hundred dollars.

18
00:01:28.390 --> 00:01:32.930
Everything stops. Permission Protocol intercepts the call and holds it.

19
00:01:33.050 --> 00:01:35.420
No named approver, no execution.

20
00:01:35.470 --> 00:01:40.810
The request lands in front of a named human. Rod signed this one.

21
00:01:41.440 --> 00:01:52.460
Every authorization mints a signed receipt, bound to the exact action a human approved.
Swap the action, and the signature breaks. Proof, not a log entry.

22
00:01:52.610 --> 00:02:05.780
And no, not every action stops here.
You draw the line: by tool, by amount, by blast radius.
Routine clears by policy, instantly.
This one crossed the line.

23
00:02:07.670 --> 00:02:12.690
The gate isn't what slows your agents down. It's the reason you can finally turn them loose.

24
00:02:12.840 --> 00:02:21.860
And when someone asks who approved this, an auditor, a board, an incident review, the answer already exists. In seconds, with proof.

25
00:02:22.010 --> 00:02:29.490
Money, deploys, migrations, credentials: routine clears by policy, consequential waits for a human.

26
00:02:29.640 --> 00:02:33.140
That list is still growing. Somebody's about to be next on it.

27
00:02:33.790 --> 00:02:35.320
It doesn't have to be you.

28
00:02:35.920 --> 00:02:39.830
Permission Protocol. The authority layer for AI agents.