How Hackers Use AI: Techniques Behind Modern Cybercrime and Defences
Hackers use AI for six main techniques: writing convincing phishing messages, impersonating people with cloned voices or deepfake video, creating synthetic identities and documents, getting help with code and scripts, attacking company AI apps through prompt injection, and poisoning data that models learn from. Each has a matching defence. Phishing-resistant MFA and reporting buttons counter AI phishing.
Topic: AI in Cybersecurity in Cybersecurity & Ethical Hacking · Explainer
Also covers: Cyber Threats, Scams & Online Safety
Quick answer: Hackers use AI for six main techniques: writing convincing phishing messages, impersonating people with cloned voices or deepfake video, creating synthetic identities and documents, getting help with code and scripts, attacking company AI apps through prompt injection, and poisoning data that models learn from. Each has a matching defence. Phishing-resistant MFA and reporting buttons counter AI phishing. Call-back rules stop impersonation. Liveness checks and document verification slow synthetic identities. Patching and EDR limit code-assisted attacks. Input isolation and least privilege reduce prompt injection, and data provenance checks reduce poisoning. Sources include MITRE ATT&CK, MITRE ATLAS and the OWASP LLM Top 10.
Key takeaways
- Every AI-assisted technique has a known counter-control.
- Process controls beat detection for deepfake fraud.
- Company AI apps are now targets in their own right.
- Prevention, detection and response all matter.
- Map controls to ATT&CK and ATLAS for coverage.
Table of Contents
- Which AI Techniques Do Hackers Use?
- Which Controls Counter Each Technique?
- How Do You Defend Against Impersonation?
- How Do You Protect Your Own AI Applications?
- How Do You Detect and Respond?
- References
Which AI Techniques Do Hackers Use?
Microsoft's February 2024 report found state actors using AI for research, scripting and social engineering. The FBI documented criminals using AI-generated text, images, voice and video for fraud. Attacks on AI systems themselves are catalogued in MITRE ATLAS and NIST AI 100-2 E2025.
Which Controls Counter Each Technique?
| Attacker technique | Prevent | Detect |
|---|---|---|
| AI phishing (ATT&CK T1566) | Phishing-resistant MFA, link isolation | User reports, mail anomaly alerts |
| Voice and video impersonation | Call-back and dual approval for payments | Unusual payment requests flagged |
| Synthetic identity and documents | Liveness checks, document verification | Duplicate device or data patterns |
| AI-assisted coding of malware | Patching, application control | EDR behaviour alerts |
| Prompt injection on company AI | Separate untrusted content, least privilege | Log and review tool calls |
| Data poisoning | Data provenance and access control | Model performance monitoring |
How Do You Defend Against Impersonation?
Do not rely on people spotting fakes. Build rules that work even when the fake is perfect:
- No payment or bank detail change on a call or message alone; call back on a known number.
- Two-person approval above a set amount.
- Code words for executives and family members.
- Phishing-resistant authenticators as described by CISA and NIST SP 800-63B.
How Do You Protect Your Own AI Applications?
The OWASP LLM Top 10 is the practical checklist.
| OWASP risk | Control |
|---|---|
| LLM01 Prompt Injection | Treat documents and web content as untrusted; require approval for sensitive actions |
| LLM02 Sensitive Information Disclosure | Limit data the model can reach; filter outputs |
| LLM05 Improper Output Handling | Validate and encode outputs before use |
| LLM06 Excessive Agency | Least privilege for tools and agents |
How Do You Detect and Respond?
Log email events, identity sign-ins, payment approvals and AI tool calls in one place, and keep them for 180 days as CERT-In requires. Rehearse a deepfake payment fraud scenario. Report qualifying incidents to CERT-In within six hours, and victims of fraud in India should call 1930. Security testing of these controls needs written authorisation under Sections 43 and 66 of the IT Act 2000. Related: social engineering attacks and prevention and Chaos GPT and malicious AI. WebAsha offers ethical hacking and SOC training.
Review the matrix at least once a year. New AI capabilities appear quickly, but most still fit one of these six techniques. When a new threat is reported, ask which row it belongs to, whether your prevent and detect controls still work, and who owns the gap. This keeps the programme focused on controls you can test rather than on headlines about new tools.
References
Frequently Asked Questions
What's Your Reaction?
Like
0
Dislike
0
Love
0
Funny
0
Wow
0
Sad
0
Angry
0