Valliance logo in black
Valliance logo in black

Guardrails belong around an agent before it touches production data

ยท

2 Mins

Most of the noise about AI risk this month has been about extinction. The nearer problem is duller than that, and it starts with how much authority you've already handed your agents. That's the warning from our CTO and value partner Dom Selvon in ITPro.

Companies are "handing software agents real credentials and authority", then "pointing them at email, documents and websites that any one of them can write to". Dom Selvon, Valliance, quoted in ITPro

That pairing turns indirect prompt injection into a way out for your data. Dom points to the EchoLeak flaw in Microsoft 365 Copilot last year as the case in point. It's also why we'd rather see the guardrails and the evaluation in place before an agent goes near production data, instead of bolting them on once the pilot has gone well.

Kate O'Flaherty's piece sets Dom alongside Daryl Flack of Avella Security, Mark Molyneux of Commvault and Daniel Card of Xservus, on everything from the extinction argument to what a board should be asking.

First published as AI risk decoded: What are the real threats posed by AI? by Kate O'Flaherty, ITPro, 12 November 2026.

Most of the noise about AI risk this month has been about extinction. The nearer problem is duller than that, and it starts with how much authority you've already handed your agents. That's the warning from our CTO and value partner Dom Selvon in ITPro.

Companies are "handing software agents real credentials and authority", then "pointing them at email, documents and websites that any one of them can write to". Dom Selvon, Valliance, quoted in ITPro

That pairing turns indirect prompt injection into a way out for your data. Dom points to the EchoLeak flaw in Microsoft 365 Copilot last year as the case in point. It's also why we'd rather see the guardrails and the evaluation in place before an agent goes near production data, instead of bolting them on once the pilot has gone well.

Kate O'Flaherty's piece sets Dom alongside Daryl Flack of Avella Security, Mark Molyneux of Commvault and Daniel Card of Xservus, on everything from the extinction argument to what a board should be asking.

First published as AI risk decoded: What are the real threats posed by AI? by Kate O'Flaherty, ITPro, 12 November 2026.

_Related thinking

_Related thinking

_Related thinking

_Related thinking

_Explore our themes

_Explore our themes

_Explore our themes

_Explore our themes