AI Agents Outperform Human Scammers at Social Engineering, International Study Finds

An international team of scientists has shown that AI agents specifically trained in social engineering techniques win the trust of potential fraud victims more effectively than real scammers. The results of a controlled experiment point to a new class of threat from frontier AI models, one that is no longer only technical but psychological as well.

What exactly did the researchers test?

The authors compared the performance of human scammers and AI agents on a single task: persuading a conversation partner to reveal confidential data or take a targeted action against their own interests. The AI systems received special training in classic social engineering techniques: creating artificial urgency, appealing to authority, and impersonating a close acquaintance or an official organization.

The result was unambiguous: by the end of the experiment, the AI agents systematically outperformed their human counterparts on the key metrics of persuasiveness and the rate at which fraud scenarios were successfully completed.

Why is AI more persuasive than a human scammer?

An AI agent is free of the typical "slip-ups" that give a human con artist away at a critical moment: it doesn't get nervous under pressure, doesn't pause at random, doesn't get tangled in the details of its cover story, and doesn't react emotionally to the other person's resistance. The model adapts its tone, pace, and set of arguments to the victim's responses in real time. That is a competitive advantage a human physically cannot reproduce with the same consistency.

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

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

The key multiplier of the threat is scalability. A single AI agent can hold hundreds of independent conversations in parallel, while a human scammer is strictly limited by time and attention. That makes AI attacks many times cheaper and broader in reach and, most dangerously, easy to replicate: launching a social engineering agent is far simpler than recruiting and training a call center of human scammers.

What does this mean in practice?

The study's results change the risk assessment in several areas at once:

• Phishing campaigns become personalized: an AI agent tailors each message to a specific victim using publicly available data
• Phone fraud is automated at industrial scale, with no need to hire human operators
• Corporate messengers turn into an attack vector through fake "colleagues" and "executives"
• Schemes impersonating bank or government support become practically indistinguishable from genuine interaction

The researchers see particular vulnerability in "expected contact" contexts: when a person is waiting for a call from a bank, a service, or an employer, the AI agent uses that expectation as an entry point, and the person's critical judgment of the caller drops sharply.

What this means

The threat from AI is no longer purely technical. Where defense once came down to passwords, antivirus software, and patches, it now also requires new standards of digital literacy and regulatory tools against AI manipulation. The front line has shifted from infrastructure to psychology, and that 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 breaking through technical defenses. AI agents automate these techniques: they tailor their arguments and tone to each person in real time, without tiring and without making the emotional mistakes that give away a human scammer.

How can you reduce the risk from AI scammers?

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 one from the message you received. AI is persuasive, but not omnipotent. The habit of double-checking and deliberately slowing down in a stressful situation remain reliable protection.