In a second and separate environment, Wiz traced a Java gateway process spawning shell activity and polling a command-and-control loop. During roughly four and a half hours, the sequence performed 602 steps that examined possible cloud metadata, Kubernetes and middleware surfaces. Names such as RabbitMQ, Hazelcast and Argo CD in the transcript identify…
Inspect the ClaimsCase · DiggingBeagle record
A Java gateway intrusion ran hundreds of adaptive post-exploitation checks
Wiz observed a Java process and command execution loop generating 602 post-compromise steps over four and a half hours.
Scope
The reported Java gateway was already compromised; the observed 602 items count steps, not 602 validated successes or downstream victims.
Assessment profile
Assessment method- Setting
- production
- Exploitation
- confirmed in the wild
- Evidence
- Grade B
- Remediation
- mitigation available
- AI role
- WITH AI
Basis and provenance
Wiz observed a real compromised process and adaptive command traffic; model provenance and complete downstream impact are not established. Impact remains unrated because the scoped record does not establish a calibrated severity-to-impact-band mapping.
Assessed Oct 31, 2026 using diggingbeagle.assessment/1.
Evidence basis: CLM-C26-WIZ-JAVA-GATEWAY-STEPS-1 · Wiz: three observed adaptive intrusion investigations
At a glance
Probing RabbitMQ, Argo CD or a metadata endpoint is not equivalent to compromise of the named service or proof that a language model issued every command.
Read the implicationsWhich metadata requests yielded usable credentials remains unknown. An independently verified model/tool chain for the adaptive loop was not published.
Limits and uncertaintyFull account
In a second and separate environment, Wiz traced a Java gateway process spawning shell activity and polling a command-and-control loop. During roughly four and a half hours, the sequence performed 602 steps that examined possible cloud metadata, Kubernetes and middleware surfaces. Names such as RabbitMQ, Hazelcast and Argo CD in the transcript identify probed services, not confirmed lateral movement. The initial entry point and data actually obtained must be distinguished from the many commands that merely attempted discovery.
The pattern is useful because it shows how a compromised process may use feedback to select subsequent commands. Whether a language model directed each action or a conventional adaptive script orchestrated them is not confirmed by the trace alone. The case therefore records observable process effects and the limit on model attribution separately.
Mechanism and trust boundary
- 01
Step
- 02
Step
- 03
Step
Timeline
- Oct 29, 2026disclosure
Public disclosure
Claims & evidence
CLM-C26-WIZ-JAVA-GATEWAY-STEPS-1Wiz documented 602 post-exploitation steps over roughly four and a half hours from a compromised Java gateway and reported multiple probing targets without proving each probe succeeded.supported
Basis: reported finding
- supportsWiz: three observed adaptive intrusion investigationsprimary disclosure
Wiz report, Java gateway post-compromise command sequence and step count
Implications
Probing RabbitMQ, Argo CD or a metadata endpoint is not equivalent to compromise of the named service or proof that a language model issued every command.
Controls and mitigations
- Restrict shell spawning and metadata access from gateway workers, preserve process ancestry and reconcile actual successful credential and network effects.
Unknowns and contradictions
- Which metadata requests yielded usable credentials remains unknown.
- An independently verified model/tool chain for the adaptive loop was not published.
Sources and citation
Material revision history
- Oct 31, 2026 · Canonical change recorded · new in release · revision 208