Attackers weaponize ARTEX AI pentest framework as LLMjacking tops $1 million in cloud charges

Weaponized AI and LLMjacking reshape the cloud threat landscape

A report published on dev.to argues that AI has crossed over from a defensive tool to an offensive one, with attackers now using automated agents to breach applications and hijack cloud resources at scale. The piece weaves together three threads: the misuse of an open-source penetration-testing framework, a wave of fraudulent cloud AI charges known as LLMjacking, and the end of "security through obscurity" as a viable strategy.

ARTEX AI moves from research tool to attack tool

According to the dev.to report, the framework at the center of several incidents is ARTEX AI, an open-source, multi-agent system built for penetration testing. Although it was designed for security research, threat actors are now using it to run automated attacks.

The report points to two concrete cases. Developer platform WorkOS confirmed a breach in which an attacker used ARTEX AI to reach an internal sales application and steal summaries of customer communications. Separately, hackers attributed to China reportedly used the same framework in automated campaigns against more than 200 organizations across 14 countries.

A third incident, a "3AM" ransomware attack on a UK school estate, is suspected but not confirmed to have been AI-assisted, the report notes.

LLMjacking: cryptojacking for the AI era

The financial dimension of the story is LLMjacking, described in the report as the AI-era counterpart of cryptojacking. Attackers steal cloud credentials or API keys and use them to run expensive large language model inference on the victim's bill.

Google has tallied more than $1 million in fraudulent cloud AI charges from such attacks, according to the report, with some organizations losing over $100,000 in a single day. The implied lesson is blunt: AI API keys deserve the same protection as root cloud credentials.

Forty years of technical debt comes due

The report also cites an IEEE Spectrum argument that AI is effectively forcing organizations to pay off "40 years of technical debt". Because agents can scan and exploit code cheaply and quickly, obscure flaws — the report names the PHP2Shell vulnerability — that might have lingered unnoticed for decades become practical targets. Neglect is no longer a defense.

That shift is expected to reshape security teams. Citing Gartner, the report predicts that by 2028 Security Operations Centers will employ 50 percent more engineers than analysts, as AI agents absorb routine alert handling and humans concentrate on engineering and remediation.

What developers can do now

The report closes with three practical recommendations:

• Rotate AI API keys for OpenAI, Anthropic and Vertex AI under strict policies, since long-lived keys compound LLMjacking losses.
• Review Gemini data settings: chat history is used for training even for personal paid subscribers. Disabling "Keep Activity" opts users out, but caps chat retention at 72 hours.
• Watch local inference performance: on Linux, some hardware configurations — the report names Nvidia RTX Spark with encrypted firmware — may see inference speeds fall to roughly 35 percent of expected performance, which can slow local security tooling.

Why it matters

LLMjacking turns AI adoption into a direct financial liability: every API key is now a meter an attacker can run. Meanwhile, ARTEX AI demonstrates that capable offensive tooling is open-source and freely available, lowering the skill floor for automated attacks on cloud infrastructure. For engineering teams, the combined effect is that long-held assumptions — that obscure bugs stay obscure, that AI keys are low-value targets — no longer hold. Treat AI credentials like production secrets, and expect security operations to reorganize around machines doing triage while humans handle remediation.