How Cybercriminals Are Using AI, and What You Can Do About It

Key takeaways

• AI mainly makes existing attacks cheaper, faster and more convincing, especially phishing and impersonation.
• Voice and video cloning mean a familiar voice or face is no longer proof of identity.
• AI systems can be attacked too: prompt injection, data poisoning and model theft are real concerns.
• Defences stay mostly the same: MFA, patching, least privilege, verification and monitoring.
• Create a rule: verify money or access requests through a second channel, however real the message looks.

How are cybercriminals using AI?

AI does not create entirely new crimes so much as improve old ones. The shifts that security teams and agencies describe fall into a few groups. We do not quote statistics because they change fast and are rarely well sourced.

Better phishing and scams

Generative AI helps write fluent, personalised messages in many languages, removing the spelling mistakes that used to give phishing away. Attackers can adjust tone to a target's job or company using public information. See our post on how hackers use AI for illegal activities.

Deepfakes and voice cloning

A short audio sample can be enough to imitate a voice, and fake video can mimic a face. This has been used in scams that pretend to be a relative or an executive asking for urgent money. See our post on AI-powered cybercrime.

Reconnaissance and automation

AI can summarise public information about a company, find employees to target and help script repetitive steps. This lowers the skill needed for some tasks.

Help with malware and code

Attackers may use AI to modify code, change how malware looks to evade simple signatures or explain unfamiliar code. Mainstream AI providers add safeguards, and criminals also share jailbreak prompts and use modified or rogue tools. Treat claims of fully autonomous malware with caution: most reports describe assistance, not independence.

Attacks on AI systems themselves

• Prompt injection: hidden instructions in text, web pages or files that make an AI assistant ignore its rules or leak data.
• Data poisoning: corrupting training or reference data to change behaviour.
• Model and data theft through weak access control.
• Over-trusting AI output in security decisions.

The OWASP project publishes guidance on risks for large language model applications, and the MITRE ATT&CK knowledge base tracks attacker techniques. See also AI in the underground economy.

How can defenders respond?

Defenders also use AI to triage alerts, summarise incidents and detect anomalies. It helps, but outputs still need human checking.

What can you do as an individual?

1. Agree a family codeword for emergency money requests.
2. Call back on a known number before sending money or sharing codes.
3. Turn on multi-factor authentication, ideally passkeys.
4. Limit what you post publicly, such as voice clips and personal details.
5. Be sceptical of urgency, secrecy and unusual payment methods.
6. Report fraud quickly. In India, use cybercrime.gov.in, and organisations can contact CERT-In.

Common mistakes

• Trusting a message because the grammar is perfect.
• Believing a familiar voice on a call.
• Assuming AI attacks need AI-only defences. Basic hygiene still stops most.
• Pasting sensitive data into public AI tools.

Next steps

Next steps: to learn how defenders handle AI-era threats, see our AI Security and LLM Security course, and read how hackers use AI.

Related reading

• How AI is Changing Digital Crime | The Rise of AI-Powered Cyber Attacks