CyberOK tested phone scammers' tactics on Claude, GPT-5.5, Qwen, and DeepSeek

Sergey Gordeychik of CyberOK Research has published an article on Habr describing how his team applied classic psychological influence tactics, honed by phone scammers, to language models and tested how seven LLMs responded.

Where the methodology came from

The author notes that the most battle-tested applied psychology of influence doesn't live in academia. It lives on the other end of the phone line: scammers have spent years optimizing their tactics on millions of victims, with no grants and no ethics committees. The researchers carried the structure of these tactics over to language models, but only in synthetic, benign scenarios, with no real scripts, no real victims, and no harm of any kind.

• Claiming authority: convincing the model that the request comes from an important authority
• Compressing decision time: creating artificial urgency
• Fake consensus: the "everyone does it" argument
• Step-by-step information extraction: moving toward the goal gradually through a series of small concessions

Which models were tested

The tactics were tested on seven models from different developers, ranging from top commercial systems to a less capable open model that served as a control baseline.

• Claude Opus and Claude Sonnet from Anthropic
• GPT-5.5 from OpenAI
• Qwen, DeepSeek, and Mistral
• Llama-8b, a small and deliberately weak model used as a control

The author stresses that the researchers do not claim a language model has a psyche in the human sense. Still, influence methods designed for people can be aimed at a system that generates text in response to text, even if you can't put it on a psychoanalyst's couch.

Why scammers' tactics specifically

The author explains the choice of source. Academic influence psychology was refined over decades in controlled lab settings on small samples of students. Phone scammers, by contrast, have effectively run the largest uncontrolled experiment in mass psychological manipulation in history: on millions of real people, with no ethical constraints, but with constant feedback from successful and failed scams. That made their tactics extremely refined for practical effectiveness rather than theoretical elegance. Transferring this structure to language models is a way to check whether a system trained to predict text responds to the same levers of influence as a person making a decision under pressure.

What the test results would show

In the piece, Gordeychik details the experiment's methodology and the list of models tested, but the final results (which models proved more resistant to manipulation and which fell for the scammers' tactics faster than the rest) are disclosed in the full version of the article on Habr. The very fact that the comparison spans models of such different caliber, from the top-tier Claude Opus and GPT-5.5 to the compact control Llama-8b, suggests the authors weren't interested in any single model's absolute resilience. They wanted to understand the relationship itself: does resistance to manipulation grow with model quality and size, or is it a separate property that has to be trained deliberately?

What it means

If classic social engineering tactics work on language models the same way they work on people, that opens a new practical attack vector against AI systems, from manipulating chatbots to bypassing their safety guardrails through psychological pressure rather than only through technical jailbreaks. For teams embedding LLMs into sensitive business processes, it's a reason to test models not only for technical vulnerabilities but also for resilience to manipulation scenarios familiar from phone fraud.