Commvault announced on November 10 a strategic partnership with Zscaler around securing AI and recovering from it when things go sideways. Standard disclosure: I work at Zscaler, so read this as informed, not neutral.
I'll be upfront about what this is and isn't. The announcement describes areas the two companies "will partner on." It doesn't describe a shipped integration or architecture yet, so I'm not going to pretend it does. What I like is the idea underneath it.
What Was Announced
Per Commvault's press release:
- Zscaler's side: Zscaler Security for AI, which discovers the AI apps in use across the enterprise and enforces policy on what users and AI agents can do.
- Commvault's side: recovering data, applications and identities after attacks, disruptions or outages.
- The three areas they named:
- Visibility and shadow AI governance: discovering the AI apps, models and tools actually in use.
- Policy enforcement and data protection: guardrails that flag unsanctioned AI activity and keep sensitive data out of places it shouldn't go.
- Agent security and recovery: securing autonomous agents in real time, and supporting clean recovery after a security event or a risky AI action.
They'll show more at Commvault's SHIFT Virtual event on November 18.
Why I Think the Pairing Makes Sense
Most AI security conversations stop at prevention. Discover the shadow AI, block the risky stuff, inspect the prompts. All good, and I spend a lot of my time on exactly that.
But agents don't just read data. They do things: change records, delete files, run workflows, call APIs with real permissions. I've written plenty here lately about agents that went off the rails, whether jailbroken by an attacker or just confidently wrong. Prevention will miss something eventually. When an agent with write access makes a bad change at machine speed, the question isn't only "how did that get through," it's "how fast can I put it back?"
The release frames it as being able to "rollback and recover from harmful agentic activity." That's the right question. Security that stops the bad action, plus resilience that undoes the one you didn't stop, is a more honest plan than pretending any single layer catches everything.
What I'd Want To See Next
Because this is an announced partnership, not a product launch, here's what I'll be watching for:
- How policy signals from the security side tie into what gets protected and recovered.
- How recovery handles changes an agent made through legitimate credentials, which are the hardest to tell apart from normal work.
- Real customer examples once it's in use.
What To Do Regardless of Vendor
- Find your agents. You can't govern or recover what you don't know exists.
- Scope their permissions like you would a new employee on day one, not like a domain admin.
- Make sure the data and systems agents can write to are actually backed up, and that you've tested restoring them.
- Log what agents do, so when you need to roll something back you know what changed.
Bottom Line
Blocking bad AI behavior is half the job. Being able to undo it is the other half. Pairing those is a sensible direction, and I'm curious to see what it looks like in practice.