F-Secure identifies 26 ways AI could harm consumers—and how to prevent these risks
The new AI harm taxonomy tracks how harm propagates through AI systems—from misleading chatbot responses to compromised autonomous agents—and identifies where defensive measures can be deployed.
Los Angeles, November 10 (Reuters) /PRNewswire/ — Artificial intelligence is now shaping how we search, communicate, make decisions, and manage our digital lives. But when it fails, how exactly does it harm people? F-Secure’s newly released AI Harms Taxonomy maps 26 distinct types of consumer harm, tracing how each one reaches users. The result is a clear picture of where protections can intervene and who is best positioned to help.
This taxonomy was developed by Dr. Megan Squire, Chief Threat Intelligence Researcher at F-Secure, to address a blind spot in discussions about AI risks. Most frameworks merely list potential points of failure but do not further connect these harms to the systems and companies capable of preventing them. F-Secure’s research goes further, mapping each layer between AI-driven threats and their impact on consumers, thereby identifying where defenses should be deployed.
Dr. Squire said: "As AI systems become more capable, the consequences of their errors may extend far beyond providing a poor answer. Those who develop these systems have a responsibility to understand how they might fail, who could be harmed when they do, and to implement safeguards before harm occurs."
This classification reflects how the nature of risk itself is changing. In the past, criminals might have targeted individuals directly, but now AI sits in between, introducing new actors and new ways for things to go wrong. An assistant acting on someone’s behalf could be manipulated into causing harm, and chatbots may confidently mislead users. At a more fundamental level, issues often begin with how AI processes instructions and memory, or which other services it connects to.
This study currently categorizes 26 types of harm into five relational types and 19 layers of intermediation, and designs 24 defensive measures to address risks at different stages. Examples include AI-generated voice and video impersonation, agents making harmful financial decisions, leakage of personal information, or acting based on outdated context. The study also examines an attack in which malicious instructions are hidden within content processed by AI systems, potentially influencing agent behavior or altering what chatbots tell users.
One of the key findings of this framework is that responsibility for preventing AI-related harm is distributed across multiple parties. Some risks can be addressed by security software on individual devices, while others must be tackled upstream by the companies developing AI products and their underlying systems—including how AI remembers information and which other services it connects to. Clarifying who is responsible for each aspect helps determine where defenses should be placed and who must build them. This also undermines the notion that individuals can remain safe simply by identifying suspicious content or repeatedly verifying information provided by AI.
As AI shifts from “generating information” to “taking action based on information,” the question of accountability becomes more urgent. A chatbot giving an incorrect answer might mislead users, but at least users remain engaged in the interaction; however, an agent with access to a user’s account and associated services could cause real consequences without such back-and-forth confirmation. Safeguards must be built around what these systems are actually doing—and this begins with understanding the difference between chatbots and agents.
Build protections for AI agents
This research is part of F-Secure’s broader effort to build trust in the AI era. Announced alongside F-Secure Trust and its strategic partnership with Quantum AI Lab QuTwo, it represents the company’s work to protect users in a world where AI agents increasingly act on people’s behalf. TrustGuard aims to help safeguard individuals during AI interactions, while TrustPath is designed to protect their digital journey from discovery to delivery. The AI Harms Taxonomy provides the research foundation for these initiatives, identifying the risks that safeguards must address—and making clear that effective protection also depends on the companies that build and operate AI systems.
Human-Centered Research for the AI Era
Digital threats today extend far beyond scams. F-Secure’s Illuminate Research team studies how people and technology interact in the real world, exploring the complex realities of trust, deception, and online resilience in the age of AI. The AI Harms Taxonomy is available on the Illuminate Research Hub.
About F-Secure
F-Secure is a people-centric, AI-driven consumer cybersecurity experience company with 38 years of experience combating digital threats. The company helps digital service providers turn trust into a high-value growth engine by protecting their customers and enabling them to lead better digital lives in a world saturated with AI-driven scams. It safeguards billions of digital interactions annually, protects tens of millions of consumers worldwide, and generates over $10 billion in value for its partners, delivering proven, large-scale impact.
Media contact:
- Meghan Sawyer, Senior Public Relations Manager, [email protected]
- Joel Latto, Threat Advisor and EMEA Public Relations Manager, [email protected]
The above information is provided by Cision: http://news.cision.com