Hackers used artificial intelligence agents to turn an exposed application endpoint into full administrative control of a server in under 24 hours.
The attackers executed hundreds of commands, stole credentials, and escalated privileges without deploying new malware or exploiting a zero-day vulnerability.
Researchers Austin Ritchie and Daxton Wirth assessed with high confidence that large language model-driven agents performed substantial portions of the intrusion.
Hackers Deploy AI Agents to Automate Server Compromise
Their strongest evidence was a live Cairn orchestration dashboard on the same IP address that initiated malicious requests. Cairn is a legitimate open-source platform for coordinating multistep agent tasks.
The attack began through an internet-facing job-submission feature on an Apache Tomcat application using Spring Batch. The endpoint accepted task definitions without authentication.

Attackers submitted jobs that invoked Nashorn, a JavaScript engine available in the application’s Java environment, allowing malicious code to run with the application account’s privileges within the existing process.
This execution initially created no child processes, limiting visibility for process-based detection. Attackers deliberately triggered exceptions and retrieved command results through error messages rather than installing a web shell.
Restricted response sizes forced retrieval in fixed 1,800-byte chunks, and sequential job names recorded file offsets and upload positions throughout execution.
Reading application configuration files exposed plaintext credentials with SQL Server sysadmin rights. The attackers enabled xp_cmdshell to execute operating-system commands under the database service account, then identified SeImpersonatePrivilege.
They staged PrintSpoofer and GodPotato, public privilege-escalation utilities that abuse that privilege to obtain SYSTEM, Windows’ highest-privilege local security account, during exploitation.
Both tools were delivered as base64-encoded fragments and reconstructed on the compromised host. After repeated file-permission failures affected the first approach, the attackers switched tools and achieved SYSTEM access.
They subsequently saved the SAM, SYSTEM, and SECURITY registry hives, enabling offline extraction of local password hashes and other stored credential material.
The command history showed rapid execution, programmatically generated identifiers, structured output, and repeated corrections addressing preceding errors. The median gap between commands was approximately six seconds.
Near-identical payloads repaired specific syntax failures, while timeout adjustments followed unsuccessful reads. These patterns suggested feedback-driven adaptation rather than a fully predetermined command sequence.
ReliaQuest cautioned that these behaviors alone do not prove LLM involvement. Investigators could not determine the number of agents, the underlying model, or the extent of human approval.
The Cairn dashboard strengthened the assessment, but investigators could not conclusively link other AI-related services found across attacker infrastructure to the operation under investigation.
Defenders should authenticate exposed management endpoints, restrict authorization, retain job histories, and correlate application exceptions with database and endpoint telemetry.
Configuration-file credentials require migration to protected secrets storage and reduced service-account privileges. Automated containment matters because early execution remained inside the application, while retained job records survived the attackers’ cleanup.
IOCs
| Indicators & Observables | Context |
| 204.194.55[.]189 | Source of requests that created malicious application jobs; also exposed a Cairn interface and exploit-staging material |
| 204.194.54[.]240 | Sent traffic to the environment prior to the incident; exposed AI-related services and an open directory containing tool-named files matching the incident |
| 94.177.131[.]113 | Retrieved job output and sent POST requests to the upload endpoint; a relay role is possible but unconfirmed |
| /tmp/out_<jobname>.txt | Command output written to a file named after the submitting job |
| /var/tmp/.x/ | Dot-prefixed directory holding shell scripts and a private runtime |
| /var/tmp/kvragent | Binary dropped on the server and used for scanning |
| C:\Windows\Temp\ps.b64 → C:\Windows\Temp\ps.exe | PrintSpoofer: uploaded as base64 fragments through the command channel and reassembled with certutil |
| C:\Windows\Temp\gp.b64 → C:\Users\Public\g.exe | GodPotato: uploaded as base64 fragments through the command channel and reassembled with certutil |
Note: IP addresses and domains are intentionally defanged (e.g., [.]) to prevent accidental resolution or hyperlinking. Re-fang only within controlled threat intelligence platforms such as MISP, VirusTotal, or your SIEM.
Join 16,000+ SOC teams using ANY.RUN to streamline threat investigations and reduce manual effort. Explore for your team