A new taxonomy has been developed by F-Secure’s Principal Threat Intelligence Researcher, Dr. Megan Squire, to map out 26 distinct harms that AI systems can cause to consumers. The AI Harms Taxonomy provides a clear view of where protections can intervene and who is best positioned to help prevent these harms from occurring.
The taxonomy tackles a blind spot in how AI risk is discussed by connecting the dots between potential harms, AI-enabled threats, and their impact on consumers. Unlike most frameworks that simply list what could go wrong, F-Secure’s research goes further by tracing every layer between an AI-enabled threat and its consequences for individuals. By pinpointing where defenses need to be put in place, the taxonomy provides a much-needed foundation for building protections against AI-related harm.
The harms mapped out in the taxonomy range from misleading chatbot answers to compromised autonomous agents, with 24 defenses designed to address risks at different points. Examples of potential harm include AI-generated voice and video impersonation, agents making financial decisions on behalf of individuals, exposing personal information, or acting on outdated context. The research also examines attacks where malicious instructions are hidden in content an AI system processes.
One of the framework’s central findings is that responsibility for preventing AI harm is spread across several parties. Some risks can be handled by security software on a person’s own device, while others have to be dealt with further upstream by companies that build and run AI systems, including how these systems handle instructions and memory, and which other services they connect to.
This raises important questions about who is responsible for preventing AI-related harm. As AI moves from producing information to acting on it, the potential consequences of its mistakes can extend far beyond a simple error. A chatbot that provides an incorrect answer may mislead someone, but at least that person is part of the exchange. An agent with access to a person’s accounts and services linked to them can cause real-world harm without any kind of back-and-forth.
To address this issue, F-Secure has developed TrustGuard, a system designed to help protect people during AI interactions, as well as TrustPath, which aims to secure digital journeys from discovery to delivery. The AI Harms Taxonomy provides a research foundation for these efforts by spelling out the risks that protections need to address.
In conclusion, the development of this taxonomy is an important step towards understanding and mitigating the harms caused by AI systems. By mapping out 26 distinct harms and pinpointing where defenses need to be put in place, it provides a much-needed framework for building protections against AI-related harm. Ultimately, meaningful safeguards depend on companies building and running AI systems taking responsibility for preventing these harms from occurring.
As consumers become increasingly reliant on AI agents to act on their behalf, it is essential that we take proactive steps to protect ourselves against potential harm. This includes being aware of the risks associated with AI interactions, understanding how our personal data is used by AI systems, and demanding greater transparency and accountability from companies building these systems. By taking a human-centered approach to researching and addressing AI-related harm, we can build trust in this technology and create a safer digital environment for everyone.
Source: Dark Reading — 2026-11-10