Cybercriminals have spent the past year adopting AI-powered penetration testing tools that security researchers say can breach companies and steal money, data and sensitive information, and those warnings now appear to be playing out in real attacks. Since around the end of August, multiple companies, government agencies and banks worldwide have been targeted by criminals using a new wave of these setups. The tools mimic legitimate defensive penetration testing techniques, using LLMs to scan targets and break into poorly secured networks before carrying out attack plans written by human operators.

Dozens of these AI penetration frameworks, or harnesses, are increasingly being linked to real-world attacks, according to alerts from around half a dozen security companies. The tools do not introduce new ways to attack websites or servers and are not sophisticated, the alerts state. Instead, they let criminals launch attacks at speeds humans could not match, and researchers say users do not need skills to operate them. Attacks that once took days now finish in hours, and CJ Moses, AWS's chief information security officer, wrote in a company blog post that agentic frameworks have moved from experimental use to sustained, multi-objective offensive operations with minimal human intervention.

The most prominent tool is Artex, an open-source framework written in Go and first published on GitHub in July by a pseudonymous developer using the handle Autumn-27. It has since spread across hacker communities and been used across all parts of attacks, from initial network analysis and gaining access to moving within systems and preparing data for export. Jacob Larsen, an offensive security researcher and founder of the firm Cypen, described it as an all-in-one product and a point-and-shoot weapon, noting earlier tools handled only reconnaissance or only testing. Artex is not itself an AI product, but a guide that points external frontier and open-source models, connected via APIs, to existing penetration testing tools and processes.

Users begin operations with simple natural-language prompts. Check Point researchers reported that one set of hackers instructed Artex to launch a penetration attack on a redacted victim to gain access to the site's admin panel, database or server. Multiple agents then break the plan into tasks and carry them out. Safety guardrails in frontier models can stop hacking activity when detected, but researchers say free Chinese and abliterated models are increasingly better at breaking into systems. Guillaume Valadon, a researcher at GitGuardian, said the tool is fast and efficient and lowers the bar, adding that his grandmother could do it except that she does not have a laptop.

Reports from cybersecurity companies have tied Artex to attempted and successful hacks, often by Chinese-speaking criminals seeking money. These include theft of personal information from thousands of South Korean financial institutions, potential cryptocurrency theft, and searches for exposed login details on GitHub. The FBI warned last week that a variety of malicious cyber actors have used Artex against organizations globally in financial services, government, technology, defense, semiconductor, energy, health care, manufacturing, telecommunications and legal industries. In October, the China-linked group Vapor Panda used it against more than 200 organizations, including think tanks, semiconductor and defense firms in the US, and European government foreign ministries, according to CrowdStrike. Microsoft observed Chinese hackers it tracks as Storm-0940 using Artex to try signing in to Microsoft 365 setups after a supply chain attack, gaining privileged access in under three hours.

On October 8, Autumn-27 removed Artex from GitHub, commenting that they are in no way associated with the malicious attacks and that the tools were built for educational and research purposes. An email to an address linked to the developer drew no response. Copies of the framework continue to be developed and used, and Valadon said GitGuardian has documented 45 AI attack frameworks, with CyberstrikeAI, Hexstrike, CAIRN and PentAGI frequently mentioned. Defenders have one advantage: Silas Cutler, principal security researcher at Censys, said the tools are not stealthy right now and generate heavy noise, such as the 80,000 requests Artex sent to agentic coding platform Dosu in 90 minutes. Basic practices such as patching software, avoiding default credentials, not exposing secrets and monitoring for unusual network activity can help, and thousands of exposed instances found by Censys and the Shadowserver Foundation include unsecured servers that let researchers watch operators at work.

More cybersecurity news from TechManNews.