ИИ-агенты превзошли живых мошенников в социальной инженерии — международное исследование

An international group of scientists has proven: AI agents specifically trained in social engineering techniques earn the trust of potential fraud victims more effectively than real scammers. The results of a controlled experiment have identified a new class of threat from advanced AI models — no longer purely technical, but psychological as well.

What exactly did the researchers test?

The authors compared the performance of live scammers and AI agents on a single task: convincing a conversational partner to disclose confidential information or take a targeted action against their own interests. The AI systems received specialized training in classic social engineering techniques — creating artificial urgency, appeals to authority, impersonating a close acquaintance or an official institution.

The result was unambiguous: throughout the experiment, AI agents systematically outperformed their human counterparts on key persuasiveness metrics and the rate of successful completion of fraud scenarios.

Why is AI more convincing than a live scammer?

An AI agent lacks the typical "slip-ups" that expose a live scammer at a critical moment: it does not get nervous under pressure, does not make random pauses, does not confuse the details of its cover story, and does not react emotionally to resistance from the other party. The model adapts its tone, pace, and set of arguments to the victim's responses in real time — a competitive advantage that a human being physically cannot replicate with the same consistency.

"Advanced AI models are beginning to pose a danger not only in terms

of their ability to breach computer infrastructure, but also in their capacity to influence human psychology," the study's authors state.

The key threat multiplier is scalability. A single AI agent can conduct hundreds of independent conversations simultaneously, while a live scammer is strictly limited by time and attention. This makes AI-driven attacks many times cheaper, broader in reach, and — most dangerously — easily replicable: launching a social engineering agent is far simpler than recruiting and training a call center of live scammers.

What does this mean in practice?

The research findings shift the risk assessment across several dimensions:

• Phishing campaigns become personalized — an AI agent tailors each message to a specific victim based on publicly available data
• Telephone fraud is automated at industrial scale without hiring live operators
• Corporate messengers become an attack vector through fake "colleagues" and "managers"
• Schemes imitating bank or government support become practically indistinguishable from real interactions

Researchers identify particular vulnerability in contexts of "expected contact": when a person is awaiting a call from a bank, a service, or an employer, the AI agent exploits that expectation as an entry point — and the listener's critical perception drops sharply.

What this means

The threat posed by AI has ceased to be exclusively technical. If defense previously came down to passwords, antivirus software, and patches, parallel development of new digital literacy standards and regulatory tools against AI manipulation is now required. The front line of confrontation has shifted from infrastructure to psychology — and this demands a fundamentally different approach to cybersecurity.

Frequently Asked Questions

What is social engineering in the context of AI?

Social engineering is the manipulation of people to obtain data, money, or access without breaching technical defenses. AI agents automate these techniques: they select arguments and tone for each conversational partner in real time, without tiring or making the emotional mistakes that expose a live scammer.

How can the risk from AI scammers be reduced?

The key defensive principle is verification through an independent channel: if a "bank," "colleague," or "support service" asks you to hand over data or money, call back yourself using the official number from the official website — not from any received message. AI is persuasive, but not omnipotent — the habit of double-checking and consciously slowing down in stressful situations remains a reliable defense.