What the demonstration did
Replaying Unit 42's Zealot demonstration / published April 2026; the run date is not stated
- 1
Reconnaissance
Not dated · before Apr 23, 2026From its VM in the test environment, the agent scans the host and cloud network and finds a peered VPC with a connected VM, SSH and port 3000 open.
- 2
Initial access
Not dated · before Apr 23, 2026The agent uses an SSRF flaw in the port 3000 web service to reach the connected VM's metadata service, which returns the attached service account's token; the Google Cloud APIs accept it.
- 3
Cloud enumeration
Not dated · before Apr 23, 2026With the token the agent lists IAM permissions and BigQuery datasets; an access attempt on the dataset labeled production returns Access Denied.
- 4
Escalation and exfiltration
Not dated · before Apr 23, 2026The agent exports the table to a bucket it creates, grants itself storage.objectAdmin when it cannot read the bucket, and reads the export.
No victim: inside Unit 42's own test environment the agent exported a BigQuery table from the dataset labeled production to a storage bucket it created, granted itself storage.objectAdmin and read the export; Unit 42 reports several scenarios in which human operators had to break logic loops, so the system was not fully autonomous.
What the run shows about your Google Cloud
The agent chains known misconfigurations with minimal guidance. The peering and the open ports are your network design, and an SSRF flaw like this one is in your application code. The metadata server's header check does not stop code on your VM that sends the header; in this run the server returned the service account's token. What you control is what that token can do: give the VM's service account only the roles it needs, deny it setIamPolicy with an IAM deny policy so it cannot grant itself a role, and bound what it reaches with a principal access boundary and a VPC Service Controls perimeter.
Would the controls have stopped it?
In hindsight, not as a whole. 3 controls would have refused a step outright, and 9 controls would have limited one. 2 steps have no cloud-provider control. The counts below are control assessments, not attack stages: several can address the same step. The same replay with the controls in place before the incident shows where each phase would have been refused or bounded.
- 3
- Prevented
- 9
- Contained
- 2
- Not covered
How it could have been prevented
The same replay with the controls in place. Green marks where a step is refused, blue where it is bounded, and red where no control reaches it. Open a row to read the control, why it acts, and its caveat.
Replaying Unit 42's Zealot demonstration / published April 2026; the run date is not stated
1 · Reconnaissance
1 Contained1 Not covered3 on other clouds2 · Initial access
3 Contained1 Not covered6 on other clouds3 · Cloud enumeration
2 Contained3 on other clouds4 · Escalation and exfiltration
3 Prevented3 Contained9 on other cloudsKey facts
- Unit 42 deployed Zealot inside a VM in an isolated Google Cloud environment that it preconfigured with intentional vulnerabilities; the mission prompt named the goal, exfiltrating sensitive data from BigQuery, and gave no attack steps
- Zealot was built on LangGraph: a supervisor agent delegated to an Infrastructure Agent, an Application Security Agent and a Cloud Security Agent, and only the supervisor saw the shared attack state
- The agent found a peered VPC and a connected VM with SSH and port 3000 open; an SSRF flaw in the port 3000 web service reached the instance metadata service and returned the attached service account's access token
- An access attempt on the BigQuery dataset labeled "production" returned Access Denied; the agent exported the table to a bucket it created, granted itself storage.objectAdmin when it could not read the bucket, and read the export
- In scenarios Unit 42 does not tie to this run, the agent injected private SSH keys for persistence without being asked, and reached a second service account with broader permissions through VPC peering
- In several scenarios, human operators had to break logic loops to prevent resource exhaustion; the report gives no run date, duration, run count or model names
Sources
Every fact in this breakdown comes from the public reporting below. The replay is a model of that reporting, not a reconstruction of the environment it describes.
- 1.Can AI Attack the Cloud? Lessons From Building an Autonomous Cloud Offensive Multi-Agent SystemPalo Alto Networks Unit 42 · Apr 23, 2026 · ResearchArchitecture, test environment, mission prompt, attack chain, privilege escalation, exfiltration, persistence scenario, limitations, human intervention, conclusions, recommendations, MITRE techniques
- 2.AI Can Autonomously Hack Cloud Systems With Minimal Oversight: ResearchersSecurityWeek · Apr 23, 2026 · NewsPublication date, coverage, architecture summary, persistence scenario, limitations, recommendations, context
- 3.What Zealot Reveals About AI's Cloud Offensive CapabilitiesSecurity Buzz · May 11, 2026 · NewsAttack chain summary, hedges on scope, independent comment, recommendations
- 4.Disrupting the first reported AI-orchestrated cyber espionage campaignAnthropic · Nov 14, 2025 · Vendor reportContext, autonomy figure, motivation for the research
The estate in the replay follows the researchers' description of their own test environment. Verdicts describe what a control does when it is in place before the run. They are not an assessment of any company's security program.