A self-directed AI malware strain dubbed ‘ClosedQuorum’ that calls commercial LLM APIs to generate its next move in real time has been reported to the security community. Because the malware relies on an LLM to produce a fresh command at every step rather than running a fixed script, traditional detection systems built on signatures, static patterns, and behavioral rules are effectively neutralized. The first line of defense against incident spread is shifting away from more sophisticated detection rules and toward outbound controls that cut off unauthorized AI API calls at the endpoint and validate the dynamic commands returned by the LLM just before execution.
Reasoning Malware: How ClosedQuorum Works
After breaching an internal network, traditional malware typically branches its behavior by receiving pre-written commands from a C2 server. ClosedQuorum takes this model one step further: it calls a commercial large language model API and uses the inference result as its next command. The attacker packs the infected host’s system information, privilege state, and network environment into a prompt, sends it to the API, and the LLM generates the next intrusion procedure on the fly.
What makes this approach dangerous is that the same sample is very unlikely to appear twice. Because every command is generated as the LLM’s inference output, there are essentially no hashes, strings, or YARA patterns for antivirus or EDR solutions to match. Communication with the LLM travels as a JSON payload over the standard HTTPS port 443, making it visually indistinguishable from ordinary cloud API traffic. Analysis shows that simple DPI or port-based blocking, and even traffic statistics, struggle to separate it from legitimate services. The same risk has already been confirmed in a case where a ‘silent intrusion’ compromised six 1st and 2nd-tier financial institutions within 72 hours, indicating that the attack surface of AI-driven intelligent intrusions is expanding across industries.
Why Traditional Detection Systems Fail
Traditional malware detection rests on three main pillars: signature and static analysis, behavior-based rules, and sandbox dynamic analysis. When an LLM generates a new command string on every call, as in ClosedQuorum, all three pillars lose their decisive clues. Even if an EDR tracks the activity of script hosts such as PowerShell, Python, curl, wget, Node.js, and rundll32, the payload content that their child processes emit changes every moment, so rule matching effectively cannot be applied.
Around the same time, BoanNews reported that approximately 34,000 customer records had been leaked from Kakao Golf Reservation. The exposed fields included names, phone numbers, product names, round dates, payment amounts, order numbers, payment and cancellation timestamps, and virtual account numbers; Kakao VX noted that card numbers and resident registration numbers were not included. Although the two incidents differ in intrusion path and motive, the fact that LLM-mediated intelligent intrusions and conventional web application data leaks now coexist within a single threat landscape has led to assessments that security operations have grown more complex by a tier. The BoanNews card-news article that first reported ClosedQuorum explicitly identified ‘blocking unauthorized AI API calls at the endpoint’ and ‘strict outbound controls’ as the first line of defense in this shifting environment.
Blocking LLM API Calls at the Endpoint: Redesigning the First Line of Defense
The starting point of the ClosedQuorum response is cutting off unauthorized LLM API calls at the endpoint. The first policy applied is process tree-based control. In the EDR, configure detection and blocking at the child-process level for script hosts such as PowerShell, Python, curl, wget, Node.js, and rundll32 when they invoke HTTPS to commercial LLM API endpoints. Because simple domain blocking is easy to bypass, it is essential to also log which parent process issued the call and with which argument string.
Host firewalls should be hardened in parallel. On Windows Defender Firewall, Linux nftables, and macOS pf, register the domains and IP ranges of major commercial LLM APIs in the block list for unauthorized processes, and allow only trusted processes on the whitelist as exceptions. It is generally recommended to pre-allow the administrator VPN range so that operator access is not severed during an emergency block.
At the network layer, SSL inspection plays a decisive role. Apply a client-certificate-based whitelist to outbound HTTPS traffic through a proxy or SSL inspection appliance, and terminate port 443 communication with unapproved endpoints at the policy level. Because LLM calls look identical to legitimate cloud traffic on the wire, system commands inside the payload (whoami, ipconfig, registry keys, Base64 payloads, Shadow Copy deletion commands, and so on) cannot be detected without SSL inspection capable of plaintext inspection.
Multi-Layered Defense: Outbound Controls and Runtime Validation
Endpoint blocking alone is not enough. Once a single endpoint is compromised, the dynamic commands returned by the LLM can be used to attempt lateral movement to other hosts in the same subnet, and the collected data can be exfiltrated disguised as another normal API call. East-west traffic monitoring and network segmentation must be designed together to prevent the chain spread of LLM-based C2.
SIEM correlation analysis should run in parallel. A rule that combines EDR script execution logs with SIEM network logs to automatically trigger an alert when a payload sent to an LLM API contains system command patterns is decisive for operational efficiency. It is recommended to enable the generative AI app control features of DLP and CASB solutions (for example, Microsoft Defender for Cloud Apps, Zscaler, and Netskope) to gain visibility into unauthorized LLM traffic and to extend inspection scope into the command patterns inside text payloads.
The last line of defense is runtime control, which validates commands from the LLM just before execution. Before script hosts run externally received code, the code must pass through a whitelist-based command filter, argument normalization, and least-privilege enforcement. The incident response playbook should add a ‘suspected AI-API-driven intrusion’ scenario and explicitly include procedures for immediately isolating infected endpoints from the network, revoking LLM API credentials, and inspecting all endpoints within the same subnet and EDR group.
In the age of AI malware, the center of gravity in security is shifting from ‘keep threats out’ to ‘keep data from leaving.’ In an environment where signatures disappear and behavioral rules miss every time, traditional controls such as endpoint and network outbound controls, SSL inspection, and runtime command validation are once again rising as the first line of defense. Because victory or defeat is determined by the gap between the speed of LLM-driven intrusion and the speed of operational controls, security organizations face the challenge of continuously updating their policies and playbooks to match the pace of LLM-based attacks.
Reference Source
This article was prepared by consulting the following original source: BoanNews — [Card News] The era of script hacking is over… the reality of autonomous intrusions bypassing defenses in real time via AI APIs
![[카드뉴스] 스크립트 해킹 끝났다… AI API로 실시간 방어망 우회하는 자율 침투의 실체](https://darkpulse.kr/wp-content/uploads/2026/10/article-job_1791-job_17-7.jpg.webp)
Leave a Reply