AI in Cyber Defence vs Offence: What Documented Reports Show
Public reports from Google's Threat Intelligence Group, OpenAI and the UK NCSC describe attackers using AI mainly to work faster at familiar tasks: researching targets, writing phishing and influence content, translating, troubleshooting code and scripting. Google reported in January 2025 that state-backed actors misusing its Gemini model had not developed novel capabilities with it.
Topic: AI in Cybersecurity in Cybersecurity & Ethical Hacking · Comparison
Also covers: Ethical Hacking
Quick answer: Public reports from Google's Threat Intelligence Group, OpenAI and the UK NCSC describe attackers using AI mainly to work faster at familiar tasks: researching targets, writing phishing and influence content, translating, troubleshooting code and scripting. Google reported in January 2025 that state-backed actors misusing its Gemini model had not developed novel capabilities with it. Defenders use AI for detection, alert triage, malware analysis and abuse disruption, including the providers themselves banning malicious accounts. Neither side has a decisive advantage. The practical lesson is that strong basics, such as phishing-resistant MFA, patching and logging, blunt most AI-assisted attacks.
Key takeaways
- Documented attacker AI use is mostly productivity, not new attack types.
- Social engineering and reconnaissance gain the most from AI.
- AI providers now publish takedown reports on malicious use.
- Defenders gain in triage speed and analysis, not certainty.
- Basic controls still stop most AI-assisted attacks.
Table of Contents
- Which Reports Document Real AI Use in Attacks?
- How Are Attackers Actually Using AI?
- How Are Defenders Using AI?
- How Do the Two Sides Compare?
- What Do These Reports Not Tell Us?
- What Should Organisations Do With This Evidence?
- References
Which Reports Document Real AI Use in Attacks?
| Source | What it covers | Headline finding |
|---|---|---|
| Google GTIG (January 2025) | State-backed actors using Gemini | Routine tasks such as research and troubleshooting; no novel capabilities seen |
| OpenAI disruption reports | Accounts banned for malicious use | Case studies of scams, influence operations and cyber activity |
| UK NCSC (January 2024) | Assessment of near-term impact | More volume and impact, mainly via social engineering and reconnaissance |
| UK NCSC follow-up | Outlook for cyber intrusion | Evolution of existing techniques rather than new attack vectors |
How Are Attackers Actually Using AI?
- Reconnaissance: researching organisations, people and public vulnerabilities.
- Social engineering: fluent phishing text and fake personas in many languages.
- Coding help: debugging scripts and translating code between languages.
- Influence operations: generating posts and articles at scale.
Google later reported, in a November 2025 update, experiments with malware that calls AI models during execution, which shows the area is changing. For the broader picture of criminal use, see how hackers use AI for cyber attacks.
How Are Defenders Using AI?
- Email and endpoint models that classify phishing and malware.
- Assistants that summarise incidents and draft investigation queries.
- Reverse engineering help for malware analysts.
- AI providers detecting and banning abusive accounts, as their reports describe.
The NCSC notes that defensive AI may offset part of the attacker gain through better detection and secure design.
How Do the Two Sides Compare?
| Factor | Attackers | Defenders |
|---|---|---|
| Main gain | Speed and scale of lures and research | Speed of triage and analysis |
| Constraint | Provider safeguards, account bans | False positives, data quality |
| Need to be right | Once | Consistently |
| Data advantage | Public information about targets | Deep telemetry from their own networks |
Our sibling article AI vs human hackers looks at the human side of this comparison.
What Do These Reports Not Tell Us?
Each provider sees only misuse of its own service. Attackers using self-hosted open models leave no such trail. The reports are also published by companies with an interest in showing their safeguards work. Treat them as useful evidence, not a complete census. Techniques against AI systems themselves are tracked separately in MITRE ATLAS.
What Should Organisations Do With This Evidence?
- Assume phishing will be well written; train staff to verify requests, not spot typos.
- Deploy phishing-resistant MFA, especially for administrators.
- Patch internet-facing systems quickly, since reconnaissance is faster.
- Secure your own AI deployments using the Guidelines for secure AI system development.
- In India, report incidents to CERT-In within 6 hours and fraud via 1930 or cybercrime.gov.in.
To build practical defensive skills, ask about cyber defence training.
References
- Google GTIG: Adversarial misuse of generative AI
- OpenAI: Disrupting malicious uses of AI
- UK NCSC: The near-term impact of AI on the cyber threat
- UK NCSC: Impact of AI on cyber threat from now to 2027
- MITRE ATLAS
- Guidelines for secure AI system development
- CERT-In
- National Cyber Crime Reporting Portal
Frequently Asked Questions
What's Your Reaction?
Like
0
Dislike
0
Love
0
Funny
0
Wow
0
Sad
0
Angry
0