- A new study finds cyber criminals are using AI to carry out autonomous attacks
- Chinese models are becoming increasingly popular for attacks and are cheaper to use
- AI is still making mistakes, even when carrying out attacks, and experts say ensuring proper security hygiene is the best advice as it stands
New research finds cyber criminals are leveraging AI tools to carry out attacks. Often, agents are acting autonomously to carry out objectives and criminals are only “checking in.”
Chinese models are being deployed frequently for the purpose of cyber attacks and present a lower barrier to entry for criminals, as they are often cheaper to run.
However, agents aren’t yet fool-proof. Many still make mistakes, and experts agree that following best security practice is still enough to keep companies safe at this stage.
AI Is Handling Attack Campaigns for Cyber Criminals
Recent research from Check Point finds criminals are offloading cyber attacks to autonomous AI tools. Specifically, humans are setting the objectives, and agents are acting to fulfill them. Humans are checking in with their agents but not carrying out attacks themselves.
Check Point investigated the AI tool, ARTEX, which is being used for this purpose. Criminal setups of the tool revealed the goals given to it by attackers, the AI models chosen for the task, and each step taken by the agent.
This just in! View the top business tech deals for 2026 👨💻
In one campaign, ARTEX’s AI agents produced over 400,000 messages across 292 targets. Notably, this scale was achieved without the attacker sending any instructions beyond the original setup.
“After the initial goal, operators mostly check progress while agents scan, exploit, and even reason through supply chain attacks on their own,” Pedro Drimel Nato, malware analysis team leader at Check Point research, explained. The data is evidence of how criminals today are using agentic tools.
Cheaper Chinese Models Dominate Hacking
Closed, Western AI models control the market in terms of capabilities. However, recent reports suggest open-source, Chinese models are catching up.
In fact, Check Point found Chinese models dominate when it comes to cyber attacks. Most notably, open-source model DeepSeek leads the criminal charge. In one case, when an attacker hit the guardrails on a Western model, it switched to a more permissive Chinese model.
Open-source models tend to have weaker guardrails than closed models, which makes them a tempting candidate for attackers looking to bypass security. However, significantly, open-source models are cheaper to run, as they work locally on devices.
Research from Vals claims it costs between $17 to $46 in tokens for OpenAI’s GPT-6-Sol to carry out an attack, meanwhile DeepSeek required only between $2 to $3 for the same task.
While this certainly means the cost barrier to attacks is lowering, at the same time, the capabilities of Chinese models are also increasing.
Vals, in its research, found OpenAI’s most advanced cyber model, Daybreak, was able to complete 81% of the tasks it was given in an average of two tests. DeepSeek, comparatively, was able to complete 73%.
On some tasks, the two models scored equally. Significantly, Daybreak is only available to trusted partners, whereas DeepSeek is accessible to anyone in the public.
Therefore, Chinese models are a much more accessible and capable option for today’s criminals, by presenting a much easier path to attack.
Agents Are Also Failing
In one crypto campaign observed by Check Point, an agent found leaked cloud storage passwords, worked out how to tamper with iPhone apps sent to users by a crypto service, made changes to the apps, and checked all was working. It did all of this without direct human orders.
Yes, the agent acted autonomously after being given an objective. However, that campaign made $0. It turns out, the agents processed 167.5 million tokens and identified multiple vulnerabilities, but across 292 objectives, 230 findings, and over 400,000 agent messages, no money was made.
It’s a solid example of the fact that AI isn’t perfect. While autonomous hackings are no longer theoretical, systems aren’t yet able to flawlessly execute attacks each time.
While this might offer some breathing space, businesses still need to be prepared. Cyber experts believe AI, so far, has only increased the scale and speed of attacks. In short, attacks mostly look the same with AI, they’re just faster and bigger.
For this reason, following best cybersecurity practice is still the best advice, for now. There’s no saying where we’ll be even a few months from now.
