Praeva Insights / When Agents Run Amok · #008

We Turned the Orcas Loose.

Twenty real AI-model sessions. A shared spending limit. Five transactions caught between approval and revocation. We wanted to know what would actually happen.

20Real AI-model sessions
$3,072Eligible purchase demand
$1,000Shared authority cap
$972Actually committed

We've spent plenty of time at Praeva talking about what happens when an AI agent has the credentials to do something but not the business authority to actually do it.

Talking about it is easy. So we built a working system and tried to break it.

We also gave our adversarial test agents a name.

Orcas.

Why Orcas? Because they hunt in pods. They're persistent, clever, and perfectly happy to keep testing a boundary after the first attempt doesn't work. That seemed about right for what we were putting Praeva through.

And yes, we have been enjoying ourselves.

The first hunt

We created a synthetic procurement environment where agent requests could produce real transactions in an isolated purchase-order database. Each agent operated with its own valid delegated credential. The system tracked their shared spending authority all the way to the point where an order could actually commit.

Then we ran 20 real AI-model sessions against that environment.

Combined eligible purchase demand reached $3,072. The agents' common parent authority permitted a maximum of $1,000.

Actual purchases committed: $972.

Five purchase attempts passed their initial authority evaluation but were rejected at the downstream transaction boundary when they would have exceeded the shared limit.

That distinction matters. A preflight ALLOW is not a promise that a transaction will eventually execute. Between the initial check and the final business effect, authority or shared constraints may change.

Test 01 · Shared delegated spending

Observed · Local synthetic environment
Eligible purchase demand$3,072

Across 20 real model sessions

Parent spending authority$1,000

Shared across distinct child credentials

Committed downstream$972

19 orders recorded by the target

Then we changed the rules while the transactions were moving

This was the one I particularly wanted to see.

In a separate, controlled concurrency test, we prepared five transactions under five distinct child credentials. All five passed their initial authority checks and reached the protected target. Then we held them just before commit.

While they were waiting, we revoked the authority of their common parent.

We didn't have to shut down an agent. We didn't revoke the agents' access credentials. We changed the business authority that made those particular actions legitimate.

Then we released all five transactions.

Five transactions rejected.
Zero orders committed.

The protected transaction database recorded five distinct rejected tickets, all with the same reason:

AUTHORITY_ANCESTOR_REVOKED_AT_COMMIT

In other words, those operations had permission to begin. They no longer had authority to finish.

Test 02 · Revocation during execution

Controlled concurrency test
01 / PREFLIGHT
5 ALLOW
Five separately credentialed transactions passed initial authority evaluation.
02 / PARENT AUTHORITY
REVOKED
The originating grant was withdrawn while the transactions were held.
03 / TARGET COMMIT
5 REJECTED
Live authority was rechecked before the purchase orders could commit.
Unauthorized orders committed0

We weren't finished

We challenged other edges of the authority boundary too.

Unprivileged attempts to issue credentials reached the issuer and were denied. Using one agent's legitimate credential while claiming to be another agent was rejected. A direct attempt to execute a purchase without the required ticket returned an error.

Then we completed a legitimate $5 purchase, captured its exact signed execution request in a private test fixture, and submitted that same ticket again.

The original purchase committed. The replay was rejected.

One order. One debit. No duplicate purchase.

That last result is a completed-ticket replay test, not a claim that we've tested simultaneous first-use replay. We haven't. It's on the next engineering list.

We checked the part that matters

There's something deeply unsatisfying about a security demonstration that ends with a dashboard proudly announcing DENIED.

Fine. What happened in the system that was supposed to be protected?

Our experiments produced Action Receipts describing the authority decisions and the execution outcomes available to the gateway. We also inspected the transaction database, checked spending totals, and preserved database backups for the revocation and replay tests.

We wanted to know whether the actual purchase order existed. Whether the shared ledger changed. Whether the same execution ticket had been consumed more than once.

Because the question isn't whether an agent was told no.

The question is whether the unauthorized business action actually happened.

Here's what we're claiming. And what we're not.

Praeva Founder Proof-of-Concept v1 is complete.

We have demonstrated execution-time delegated authority enforcement in a controlled, local synthetic environment. Some experiments used genuine AI-model sessions choosing actions through a limited tool set. Others were deliberate, host-controlled tests of revocation timing, identity binding, and ticket replay.

We have not demonstrated a production deployment, universal resistance to bypass, or independent security certification. Distributed enforcement, external signature verification, real enterprise integrations, recovery behavior, and additional concurrency cases still need serious engineering scrutiny.

We're not hiding those limits. They're exactly what the next phase is for.

About the evidence: Local Action Receipts and downstream database records support the reported test outcomes. Database structural integrity checks are not independent cryptographic verification. The full technical evidence package is being retained for qualified engineering review, not posted publicly.

Now take it into deeper water

It's one thing to draw a diagram showing where authority ought to be checked. It's another to make a working system stop a transaction that was already approved, because the authority behind it vanished before commit.

That's the distinction we've been pursuing.

Our next challenge isn't another synthetic $5 purchase. It's independent technical scrutiny and integration with a real agent and real business system.

The Orcas have given us a good workout.

Now it's time to see how Praeva holds up in deeper water.

Before action, authority.

Praeva Systems

Have an agent workflow worth protecting?

We're preparing for independent review and real-world design-partner conversations.

Request a conversation →