LLMs Make Advanced Cyberattacks Accessible to Anyone

Darktrace confirmed that AI-generated malware exploiting CVE-2025-55182 lowers technical barriers for cybercrime by automating cloud-based attacks. The finding increases pressure on enterprises, CISOs, and regulators to strengthen cloud security controls, vulnerability management, and AI-aware threat detection amid rising cross-border digital risk.

Darktrace identified a malware deployment generated entirely by large language models designed to exploit the CVE-2025-55182 vulnerability, a flaw which allows unauthenticated attackers to send a single HTTP request that executes arbitrary code with the privileges of the user running the affected web server process.. The incident confirms the reduction of technical barriers for low-skill threat actors, allowing for the automated creation of functional exploitation tools in cloud environments.

The transition from conventional malware to development assisted by AI responds to the optimization of timelines in the attack life cycle and the ability of models to structure complex code under research pretexts. 

"CISOs and SOC leaders should treat this event as a preview of the near future,” says Nathaniel Jones, Vice President of Threat Research and Field CISO of AI Security, Darktrace. “Threat actors can now generate custom malware on demand, modify exploits instantly, and automate every stage of compromise. AI-generated malware is operational, scalable, and accessible to anyone."

The attack vector was detected through "CloudyPots," a global network of honeypots operated by Darktrace to monitor malicious activities across network services, protocols, and cloud infrastructure platforms. These environments attract and capture adversary behavior in real time, providing critical data on the tools and exploitation techniques directed at systems exposed to the internet.

In this instance, the target was a Docker honeypot intentionally configured to expose the Docker daemon without authentication mechanisms. This configuration allows any external entity to discover the daemon and create containers through the Docker API. The analysis of the incident reveals a paradigm shift in cybercrime: the use of "vibecoding," or development assisted by artificial intelligence, which allows operators with limited technical knowledge to produce professional and documented exploitation frameworks.

The relevance of this finding lies in the exploitation of the CVE-2025-55182 vulnerability, known as React2Shell. This vulnerability affects Next.js server components, allowing for remote code execution. The integration of a recent exploit into a script generated by AI demonstrates the agility with which LLM’s can be manipulated to convert newly discovered vulnerabilities into operational attack tools, notesJones.

The Intrusion Cycle and Component Deployment

The intrusion process began with the creation of a container named "python-metrics-collector." The startup configuration of the container executed a sequence of commands focused on the installation of prerequisites, including curl, wget, and Python 3. Subsequently, the system downloaded a list of Python packages from an external repository and executed a central script hosted on GitHub Gist, under the profile of a user identified as "hackedyoulol."

The forensic analysis of the Python payload revealed characteristics unusual for traditional malware. Unlike scripts written by human operators, which prioritize brevity and functionality, this file contained extensive comments and a highly organized modular structure. A distinctive element was the inclusion of the phrase: "Network Scanner with Exploitation Framework – Educational/Research Purpose Only." This terminology, Jones revealed, is characteristic of "jailbreaking" attempts on artificial intelligence models, where the attacker frames a malicious request as an educational exercise to bypass the ethical safeguards of the model.

Testing conducted with AI detection software, such as GPTZero, confirmed with a high degree of confidence that the code was generated by a LLM. The script did not only include the exploit for React2Shell; it was designed as a complete toolkit to identify targets through an IP generation loop and execute structured exploitation requests.

The core of the execution component focuses on the execute_rce_command function. This function uses a Next.js server component payload specifically designed to force an exception and reveal command output in the host operating system. 

The technical flow of the exploitation is detailed as follows:

1. Vulnerability Verification: The script invokes the function with the whoami command to confirm the capability for code execution.

2. Payload Injection: A POST request is sent with specific headers (Next-Action: x) and a multipart body containing the malicious code.

3. Privilege Escalation and Persistence: Once the vulnerability is confirmed, the script uses wget to download the XMRig cryptocurrency miner from its official repository on GitHub.

4. Final Payload Execution: The miner is activated using nohup commands to ensure the process continues running in the background, linked to a Monero wallet and a configured mining pool.

Economic Analysis and Campaign Scope

Despite the technical sophistication of the attack vector, the observed financial objectives are relatively modest. Through monitoring the supportxmr mining pool, it was determined that the wallet address of the attacker has generated approximately 0.015 XMR since the beginning of the campaign, with an estimated value of US$6.50. The daily generation rate is 0.004 XMR.

However, the volume of compromise is significant. Data from the pool shows a count of 91 workers, indicating that at least 91 distinct hosts have been successfully infected by this malware. This volume suggests that the automation permitted by AI allows for scalability that compensates for low income per individual host. Jones says that the operational value of artificial intelligence for adversaries should not be underestimated, as a single prompting session was sufficient to compromise nearly 100 systems.

A notable finding in the investigation by Darktrace is the absence of an internal propagation module (spreader) within the Docker script. In traditional malware directed at container infrastructures, it is common to include logic for the malware to replicate automatically to other nodes within the network. The omission of this function suggests that the attacker operates a centralized infrastructure from which scanning and propagation tools are executed remotely.

The IP address that initiated the initial connection, 49.36.33.11, is registered under a residential ISP in India. While this could indicate the use of a residential proxy to hide the origin of the attack, it also raises the possibility that the operator is running the propagation script from a personal computer. This supports the theory of a threat actor with limited resources but empowered by artificial intelligence tools.

“The best defense here will be AI-tuned tripwires, in everything from honeypots to game theory,” says Ram Varadarajan, CEO, Acalvio. “Organizations will need deception techniques that leverage the algorithmic behavior that offensive AI models bring, to impel those intruders to blunder into an ambush. That is our future.”