Can you trust an AI system to police another AI system? Former red teamer and Assury Founder David Girvin says absolutely not. Join us for the next episode of #RogueAgents, where Druva's VP of Product Gen AI, David Gildea sits down with Girvin to tackle an industry-wide security problem: why AI can't govern AI. We’re diving straight into the hard truths, including: 🔴 Why agents need architectural isolation 🔴 The danger of giving agents their own credentials or JIT tokens 🔴 How to actually prove what an agent did (and why) after the fact Join the conversation 👇
Rogue Agents: Why AI Can't Govern AI — with David Girvin, Founder of Assury
www.linkedin.com
An AI grading another AI's homework isn't oversight. It's like two students copying off each other's answers during an exam. Real governance means credentials the agent can't escalate on its own, isolation it can't quietly route around, and logs it genuinely can't rewrite. If your control mechanism runs on the same substrate as the thing it's supposed to control, it was never actually a control mechanism.
If the governance AI itself is manipulated, it may approve malicious actions. The governor becomes compromised.
Hallucinations
As my fellow engineers here have pointed out: AI Guardrails protecting another AI, both probabilistic and sustained by semantic prompts... does that really work??? This week we saw the case of an OpenAI agent that simply did what it thought was "correct." This illustrates how serious things have gotten. Many people, including companies billing billions from all this, hide the fact that they don't have a deterministic "brake" for this. And they don't care because only the cash matters...