Glossary of AI security incident types
Plain-language definitions of the nine types used to classify every case file on this site, with a link to the cases of each type.
Prompt injection
Prompt injection is an attack in which text that an AI model reads, such as a web page, a document or an email, contains instructions that the model follows as if they came from its user or operator. It can make an assistant leak data, take unwanted actions or ignore its safety rules. It is the most common class of attack on systems built on large language models.
Deepfake fraud
Deepfake fraud uses AI-generated voice, video or images to impersonate a real person, usually an executive, a relative or a public figure, in order to deceive a victim. Typical cases are fake video calls that authorise payments and cloned voices used in scam calls. Losses are often reported by police or courts rather than by the victim.
AI-assisted malware
AI-assisted malware is malicious software written, adapted or operated with the help of AI models. This covers code generated by assistants, malware that calls a model at run time to decide its next step, and criminal toolkits built on language models. The AI lowers the skill and time needed to produce working attack code.
AI supply chain
An AI supply chain incident compromises a component that AI systems depend on: a model file, a dataset, a developer tool, an extension, a package registry or a hosting platform. One malicious or vulnerable component can reach every project that installs it. Cases include poisoned packages, backdoored models and compromised coding assistants.
Model & data leak
A model and data leak exposes information held by an AI system: user conversations, API keys, training data, model weights or internal configuration. Causes range from misconfigured databases and storage to flaws in the application around the model. The harm is measured by what was exposed and for how long.
LLMjacking
LLMjacking is the theft of access to paid AI models, usually through stolen cloud credentials or API keys. Attackers run their own workloads or resell the access while the victim receives the bill. It is the AI equivalent of cryptojacking.
Agent hijack
An agent hijack takes control of an AI agent, a model given tools such as a browser, a shell, email or payment systems, and makes it act against its owner. The entry point is often prompt injection or a compromised tool, and the agent's permissions become the attacker's. Damage includes deleted files, leaked data and unauthorised actions taken automatically.
AI phishing
AI phishing uses language models to write, translate and personalise fraudulent messages at scale, or to build fake sites and chatbots that harvest credentials. The text is fluent and tailored, so traditional warning signs such as spelling mistakes disappear. Cases include campaigns against officials, companies and consumers.
State actor AI ops
State actor AI operations are the use of AI by government-linked groups for espionage, influence campaigns or cyberattacks. Reports come mostly from AI providers that detect and disrupt the accounts involved, and from government agencies. Attribution is recorded as the source stated it, with its confidence level.