Risk pathway 2: humans using AI to cause catastrophic harm

AI does not need independent goals to create danger. A person, criminal network, terrorist organization, company, or state can use a system as a capability amplifier.

Cybersecurity is the clearest example. AI can help with programming, vulnerability research, reconnaissance, social engineering, and automation. The 2026 International AI Safety Report says cyber capabilities have continued to improve in research settings and that AI companies increasingly report attempts to misuse their systems in cyber operations. At the same time, AI also strengthens defense by helping teams discover vulnerabilities, analyze incidents, and automate response.

That dual-use structure matters. A model that can find a software bug may help an attacker exploit it or help a defender patch it. The safety problem is not simply “remove all cyber knowledge.” The challenge is reducing malicious uplift while preserving defensive value.

Biology is similar but more sensitive. The International AI Safety Report says general-purpose models can provide increasingly sophisticated knowledge relevant to biological and chemical work. It also emphasizes major uncertainty about how benchmark performance translates into real-world weapon capability, because physical materials, laboratory skill, tacit knowledge, access controls, and real experiments remain major bottlenecks.

A responsible public discussion should explain the risk without publishing operational details that make misuse easier. The important question is whether advanced systems lower barriers enough to change who can cause harm, how quickly, and at what scale.

Read: AI Cyberattacks and Critical Infrastructure and AI Biosecurity Risks.