@irelia0nerf3Our AI agent passed the eval. But was it authorized to act?
I’ve been thinking about a gap in how we evaluate AI agents.
We test whether they can reason correctly, call the right tools, and complete tasks.
But what happens when an agent makes the right decision without actually having the authority to execute it?
I’m Alex, founder of FoundLab, a small engineering team in Brazil working on REX Guard.
We’re exploring how to separate AI reasoning from institutional authorization, enforce permissions at execution time, and produce independently verifiable evidence of what happened.
We’ve published our reference architecture and built a banking demonstration using Google Cloud and Vertex AI. The demo uses Bradesco solely as an illustrative scenario and is not affiliated with, endorsed by, or deployed by the bank.
We’re now preparing for more demanding evaluations and potential Tier 1 financial institution pilots.
And I’d genuinely appreciate some outside perspective before we get there.
If you were testing this system, what would you try to break first?
Replay attacks? Revoked permissions? Concurrent transactions? Prompt injection reaching privileged tools? A timeout where the payment actually succeeded?
Especially interested in hearing from people operating real workloads on GCP.
We have a lot left to validate. Even one difficult test case or an uncomfortable architectural question would be valuable.
Public architecture:
github.com/irelia0nerf/rex-guard-architecture
Project and demo:
ai.lounge.ai/p/shipped-rex-guard-verifiable-controls-for-ai-agents-nov-3-2026
Not here to sell anything. Just looking for thoughtful engineers, honest feedback, and perhaps a few people interested in testing this with us.
What would you test first?
Comments1
The timeout where the payment actually went through is the test I'd run first, because that's the one a bank's customers would feel right away: money sent twice, or someone told a payment failed when it didn't. Proving the agent was allowed to act and that each payment only happens once is what a Tier 1 bank will need to see before it trusts REX Guard with its customers' money.
1 comment
The timeout where the payment actually went through is the test I'd run first, because that's the one a bank's customers would feel right away: money sent twice, or someone told a payment failed when it didn't. Proving the agent was allowed to act and that each payment only happens once is what a Tier 1 bank will need to see before it trusts REX Guard with its customers' money.