Counting before claiming

Our security research team built a dataset of AI incidents against real victims – every case checked, every claim graded by evidence. They pulled apart three things the headlines keep blurring: criminals using AI as a tool, attacks run by AI itself, and AI experiments that broke out of the lab and reached real systems.

How we counted

The chapter rests on a dataset of AI-related incidents collected from the cyber security industry, AI companies and the press. Every incident is classified in three independent ways, and every entry is graded by how well it is evidenced – so a vendor’s claim and a confirmed case are never counted as the same thing. The chapter also states the limits of the dataset openly.

Take it to your team

• Who on our team decides whether an incident was really AI-driven?
• Which AI tools inside our own company could an attacker turn against us?
• Are we preparing for the AI threat in the headlines – or the one in our logs?
• What is our rule for AI agents that have access to company systems?
• Could we spot an attack that is largely run by a machine – at machine speed?

Explore some other questions:

• Can a fake protest movement target your brand? → Explore | Cognitive

• Your biggest insider risk – malicious or well-meaning? → Explore | Threat Detection

• What do takedowns really do to ransomware gangs? → Explore | Law Enforcement

• How many extortion gangs are active – and who leads? → Explore | Cy-X

• The CVE flood: did it reach real networks? → Explore | Vulnerabilities