Guarding Agentic AI
"Agentic AI is about to recreate the identity sprawl problem a hundredfold"
Nir Mishal, CISO at Orca Security, joined CTech to share his thoughts on agentic AI security, and why identity sprawl, not a rogue model, is the failure mode he’s most worried about.
“I fear identity sprawl at machine speed. The industry spent a decade cleaning up long-lived credentials and over-privileged service accounts, and agentic AI is about to recreate that problem a hundredfold, with actors that operate faster than any human review cycle,” said Nir Mishal, CISO at Orca Security, on his biggest fear about agentic AI's rollout across the industry.
“My fear isn't a malicious model; it's thousands of well-intentioned agents with standing privileges, chained tools, and no owner, in organizations that adopted them faster than they inventoried them,” he explained. “When something goes wrong, it won't go wrong slowly. The gap between adoption speed and governance maturity is the single biggest risk I see in the industry right now, and it's widening.”
CTech reached out to a spread of Israeli companies to find out how they're actually handling agentic AI security, and whether local security leaders are ahead of the curve on the risk, or simply closer to it.
Are any AI agents currently operating with real autonomy?
Honestly? Nothing in our environment runs fully unsupervised, and that's deliberate. We treat autonomy as something an agent earns per task, not a default setting. Agents today run autonomously in narrow, read-only scopes, enriching alerts, triaging findings, drafting responses, but any action that changes state, touches production, or moves data still requires a human approval gate. The technology could do more; our governance intentionally lags the capability. In my view, the CISOs who will get burned are those who let autonomy expand by drift rather than by decision. We published a company-wide AI acceptable-use policy early, precisely so the boundary is explicit rather than something each team improvises.
What security controls are in place versus on the roadmap?
Our approach starts with identity: every agent gets its own scoped identity, never a shared human credential, with least-privilege permissions and short-lived tokens. Agent actions are logged and attributable, AI tools and integrations go through security review before approval, and every agentic workflow has a defined kill switch. Revoke the identity, and it stops. Where we keep investing: ever-deeper visibility into which agents exist, what tools they can call, and what their blast radius would be, alongside detection that flags an agent acting out of character, not just out of permission. The identity layer came first because that's where agentic risk actually lives.
Have you had an incident or near-miss?
Nothing I'd classify as an incident. But I'll be honest about the pattern every CISO should watch for: over-permissioning. AI tools and agents routinely request far broader access than the task requires, and [permissions] that nobody would blink at for a human tool become dangerous with an autonomous actor behind them. That's exactly why review gates come before access is ever granted. The goal is that a near-miss dies as a rejected permission request, not as a forensics exercise. Agentic failures look boring: not a rogue AI, just a permission that was too wide. Our controls center on identity and access for precisely that reason.
Where has AI already made things better or safer?
The clearest win is speed on the defensive side. AI-assisted triage now handles the enrichment and first-pass analysis that used to consume hours of analyst time, so my team spends its attention on judgment calls instead of data gathering. Security questionnaires and compliance evidence gathering, a huge hidden tax on any security team, have been dramatically accelerated. And as a cloud security company, we use our own AI-driven platform daily, which keeps us honest: we experience both sides of the agentic equation. Net effect: the same team covers meaningfully more ground. AI hasn't replaced anyone here. It's removed the work nobody should have been doing manually.
