The Vercel incident was very simple. An employee connected an external AI tool to the company's Google Workspace through OAuth, and an attacker took over the account. From there, he had access to any internal data not marked as sensitive.
Clearly, this is not a real hack or a notorious zero-day exploitation. Eventually, someone, legitimately, connected an AI tool and clicked "Allow". That's it. The game is over.
I mean, let's be honest. Everyone at least once, connected a new AI tool, downloaded an Add-on, or used an unknown app "just to try it", didn't you? And in many cases, we linked it to our Google Workspace or GitHub. Did anyone connect it to production? Raise your hand.
How do you keep your environment trustworthy?
Not sure? Here are my most effective controls and routines to avoid mistakes and trust related risks:
- Implement multi-factor authentication (MFA). It's not a "V" for auditors. It's a mandatory measure against identity theft. Enforce it on all identities.
- Map and document your authentication and authorization methods and apps. Having a list in your head is the most common mistake. You will be surprised to realize how many forgotten connections you may miss.
- Delete identities without a clear justification. "Not sure about this identity" is not a use case, it's a vulnerability.
- Review your apps and permissions frequently. Everything with access to sensitive data and/or central apps like your Email, Drive, or Repo, should be constantly reviewed. Although it is boring as hell, re-validate why it needs access and can access be reduced.
- Cancel the connection you made for free tools. With all due respect, business innovation is significant, but it is not worth uncontrolled access.
- Rotate everything that could leak. API keys, tokens, database credentials, everything! Do not wait for signs of a leak. Just rotate.
- Analyze your logs. Seek out actions outside of your day-to-day patterns.
- Separate live accounts from testing accounts. AI tools testing should not include production data. Such separation will eliminate human errors.
- Wherever possible, mark secrets as sensitive and obviously, deny access to sensitive data.
- Simulate an incident. Knowing what needs to be done, by whom and how long it should take for whatever is important to the organization will increase your ability to recover. Have a documented playbook if possible.
With AI evolving, these controls are no longer an edge case.
The entire AI revolution relies on countless numbers of tools with connectivity and broad access. The downside is that we help attackers by making their attack path easier. Once they are in, everything is accessible, making the above recommendations the norm and even the default.
To conclude:
If you want to ensure the trustworthiness of your environment, you need to review and control the “mess” that AI tools create.
You probably cannot avoid it, but if you oversee and govern it, you can keep it trustworthy.
Not sure how to start maintaining your environment trustworthy?
Contact us and we will walk you through it.