AI risks
The harms AI systems cause in normal operation: bias, deepfakes, identity fraud, privacy leakage and misalignment. Distinct from adversarial AI attacks.
The top 8 AI risks
The eight AI risks that matter most: bias, data poisoning, privacy leakage, model theft, hallucination, opacity, misuse and supply chain exposure.
What is bias in AI systems?
Bias in AI systems comes from training data, labelling choices and proxy variables, and shows up as systematically different outcomes between groups.
AI identity fraud
How generative AI changed identity fraud: synthetic identity documents, deepfake video calls, cloned voices defeating authentication, and what still works.
An AI risk is a way an AI system can cause harm in normal operation, with nobody attacking it. That makes risk a different subject from AI attacks, where an adversary is deliberately trying to break the system. A model that quietly discriminates against applicants is a risk. A model manipulated through prompt injection is an attack. Most real programs have to handle both.
This hub covers the harms themselves. For the controls and evidence that management and regulators expect against them, see AI Governance.
Start here
- The top 8 AI risks — bias, poisoning, privacy leakage, model theft, hallucination, opacity, misuse and supply chain exposure.
- Managing AI risks: practices and resources
- What is AI Safety? — and why it is a different field to AI security.
- AI oversight at board level — what directors need to ask and see.
Harms to people
The risks that land on individuals rather than on systems. These are also the ones regulators move on first.
- What is bias in AI systems?
- AI identity fraud — synthetic identities, deepfake video calls and cloned voices defeating authentication.
- Deepfake videos: risks to privacy and prevention
- AI in physical security and surveillance
- AI security cameras
- What is the risk of autonomous weapons?
When the model itself is the risk
Failure modes that come from how models are built and trained, not from misuse. Deceptive alignment — models that behave under test and differently in deployment — is covered in AI safety, along with how red teaming actually works.
- The alignment problem — why intended goals are hard to specify.
- An AI agent exploited an API nobody asked it to touch — what happens when an autonomous agent has more agency than the task required.
- Human in the loop: where it helps and where it does not
- What is AI detection?
Data and privacy
- What is AI privacy?
- Shared AI chats indexed by search engines — a sharing feature working as designed, exposing CVs and unpublished work.
- Chatbots tricked into leaking personal data
- Data anonymization in AI systems
Training data can also be extracted from a deployed model directly — see AI attacks for how that works and why it is a data-protection problem rather than only a technical one.
Where risk becomes misuse
The boundary cases, where a capability is turned to harm. Past this line the subject becomes AI attacks.
- What is WormGPT?
- LLM agents that can autonomously hack websites
- Will AI replace cyber security jobs?
Managing the risk
Naming a risk is the easy half. Assigning an owner, assessing it against a framework and holding evidence an auditor accepts is the work — that is what AI governance covers, and the NIST AI RMF is the usual starting structure. AISGRC provides a free scorecard as a baseline for that work.