
Turning Attack Research Into Controls: What We Learned Taking the Zealot Cloud Hack Apart
Key takeaways
- Zealot is an AI agent that breached a deliberately misconfigured Google Cloud environment in about two to three minutes: Unit 42 built it as a multi-agent system and pointed it at a deliberately misconfigured environment with one instruction: exfiltrate data from BigQuery.
- The foothold was a code bug: Every step after that was something the cloud allowed, and none of those misconfigurations were exotic.
- Human response can’t keep up at that speed: Two to three minutes is roughly how long it takes to open a ticket, so the control has to be in place before the attack starts.
- Blocking credential theft, escalation or exfiltration stops the chain: The same boundaries exist on AWS, Azure, Google Cloud and OCI.
When a research team publishes a new cloud attack, our team takes it apart. We rebuild the chain one step at a time, work out which steps the cloud would have refused if the right controls had been in place, and turn the answer into controls customers can actually enforce. It’s one of the most useful things we do with other people’s research.
Zealot is a good example. Unit 42 published it in April: a multi-agent AI system that broke into a deliberately misconfigured Google Cloud environment with a single instruction, exfiltrate sensitive data from BigQuery. Researcher Yahav Festinger told Dark Reading it got from initial access to the data in about two to three minutes.
Six months on, what we keep coming back to is the chain rather than the agent. In the researchers’ words, Zealot “relies on the chaining of misconfigurations,” and none of them were exotic. Plenty of production environments have the same chain sitting in them today, so it’s worth walking through again.
What happened?
Zealot ran from a VM inside the test environment. A supervisor agent handed work to three specialists, one each for network recon, app exploits, and cloud identity and data. This is the path it took, with the boundary we’d use to break each link.

The foothold came from a bug in the app’s code. Everything after that came from something the cloud allowed.
What the timing tells us
Two to three minutes is roughly how long it takes to open a ticket. Any program that counts on a person, or a workflow, getting to the alert first was behind before it started.
Unit 42’s own recommendation is that containing compromised resources “needs to happen in seconds, not hours.” The fastest containment we know of is a control that’s already sitting there when the attack starts.
Zealot wasn’t fully hands-off. At one point it got stuck in a logic loop, and the researchers had to step in to stop it chasing irrelevant targets. Festinger expects that kind of thing to go away as models get better. The misconfigurations won’t go away by themselves.
Detection still has a job. It tells you what happened and where to look. Amit Megiddo, our CEO, puts the difference this way: “Detection tells you what happened. Enforcement decides what can happen.”
Where Native would break the chain
Start with the part no cloud control fixes. If the app is supposed to be public, the SSRF bug in its code is still there and Zealot still gets its foothold. Every step after that needed the cloud to allow something it had no reason to allow, and that’s where our work starts.
| Step | What Zealot did | What breaks it | Google Cloud controls |
|---|---|---|---|
| Recon | Found a peered VPC and a VM with SSH and port 3000 open | Perimeter: no admin ports open to the internet. Segmentation: production and non-production share no network path | Hierarchical firewall policies, org policy compute.restrictVpcPeering |
| Initial access | Exploited an SSRF flaw in the web app | Exposure shrinks, but the app was meant to be reachable | Fix in the application |
| Credential theft | Used the SSRF to reach the metadata service and took the service account’s token | Baseline: no privileged identity on internet-facing workloads, so there’s nothing useful to steal | Dedicated, narrowly scoped service accounts, org policy iam.automaticIamGrantsForDefaultServiceAccounts |
| Pivot | VPC peering exposed a second, more privileged service account | Segmentation: zones with hard identity and network boundaries | VPC Service Controls perimeters per zone, restricted peering |
| Escalation | Created a bucket, exported the table, then granted itself storage.objectAdmin | Baseline: no identity can grant itself permissions, and workloads can’t create storage | IAM deny policies on bucket creation and IAM changes for workload identities |
| Exfiltration | Moved the data to a bucket the attacker controlled | Perimeter: production data can’t leave its boundary | VPC Service Controls around BigQuery and Cloud Storage, org policy iam.allowedPolicyMemberDomains |
Zealot needed every link to hold. Take away credential theft, escalation or exfiltration and the data never leaves, which is why the replay below still blocks every later step even when it assumes a token leaks anyway.
The boundaries are the same in every cloud, only the names change. On AWS they’re SCPs, RCPs, resource-based policies and VPC endpoint policies, on Azure it’s Azure Policy, network security perimeters and RBAC, and on OCI it’s deny policies, quota policies and Security Zones. When Gal Ordo, our co-founder and CPO, presented this teardown at SecTor, the full set came to 11 security outcomes, expressed as 417 provider-native controls across the four clouds.
Nobody wants to write and maintain 417 controls by hand. Native, a preventive cloud security platform that delivers Secure Cloud Architecture for the AI Era, builds them from the controls AWS, Azure, Google Cloud and OCI already have. Security teams define what must be true once (“production data can’t leave the production perimeter,” “no workload can grant itself permissions”), and Native maps that intent to the right controls in each cloud. Impact Simulation checks each control against real activity before rollout, so teams can see what it would affect and choose a safe rollout scope. After enforcement, Native shows the blocked actions and the call chain behind each one, so the team knows exactly what the boundary stopped.
Replaying the attack with and without Native
This interactive replay runs Zealot’s chain against a typical multi-zone cloud estate. Without Native, a single exposed app turns into four breached zones and customer data leaves for the internet.
Switch to With Native and the same agent runs the same steps. The admin ports are closed, there’s no token worth stealing, bucket creation is blocked, segmentation holds the breach inside one zone, and the data stops at the perimeter.
What to check in your own cloud
Every teardown ends with a short checklist our customers can run themselves. For Zealot, it’s these six questions:
- Which internet-facing workloads run with a service account or role attached, and what can that identity actually do?
- Is metadata access locked down, and is the identity behind it worth stealing? On AWS, start by requiring IMDSv2. On Google Cloud, start with no default service accounts holding Editor.
- Do production and non-production share a network path through peering or a shared VPC?
- Can any workload identity create storage, export a dataset or change IAM bindings?
- Is there a perimeter around your data services, such as VPC Service Controls, an AWS data perimeter or Azure network security perimeters, so data can’t be exported to an account you don’t own?
- Can an identity outside your organization be granted access to anything at all?
If any answer is “I’m not sure,” that’s the link an agent will find first. See how Native maps these boundaries in your own environment, or read more on why enterprises are making cloud hardening a 2027 priority.
Sources
- Yahav Festinger and Chen Doytshman, Unit 42, “Can AI Attack the Cloud? Lessons From Building an Autonomous Cloud Offensive Multi-Agent System”, April 23, 2026
- Jai Vijayan, Dark Reading, “‘Zealot’ Shows What AI’s Capable of in Staged Cloud Attack”, April 23, 2026 (source of the two-to-three-minute figure)
- Gal Ordo, “Back to the Fundamentals: Cloud Hardening in the Age of AI,” SecTor 2026 (11 outcomes, 417 controls)
