AI Cybersecurity Incident
An AI cybersecurity incident is any event where an AI system is compromised, manipulated, or causes unintended harm with security consequences. That covers a wide range of outcomes: a language model manipulated through prompt injection to exfiltrate data, a malware classifier evaded through adversarial inputs, a fine-tuned model leaking sensitive training data, or an autonomous agent taking harmful actions it was not authorized to take.
What makes an AI incident distinct is not the presence of AI in the environment. It is when the AI system itself is the vector, the target, or the cause.
How AI Incidents Differ From Conventional Incidents
Conventional security incidents follow a familiar pattern. A vulnerability is exploited, a system is compromised, indicators of compromise appear in logs, and responders work backward from evidence to reconstruct the attack.
AI incidents break several of those assumptions.
The attack surface is the input. Prompt injection, adversarial evasion, and data poisoning all operate through inputs the system was designed to accept. There is no exploit in the traditional sense—no CVE. No patch. A malicious prompt is syntactically indistinguishable from a legitimate one.
The failure can be silent. A poisoned classifier fails quietly, returning confident but wrong outputs with no error state, no crash, and no alert. An agent manipulated via indirect prompt injection may take harmful actions that appear as normal system activity in the logs.
Causation is difficult to establish. AI models are probabilistic and opaque. Attributing harmful output to manipulation rather than to a model limitation, a hallucination, or an unusual input distribution requires forensic capabilities that most organizations have not built.
The blast radius scales with autonomy. An AI agent with broad access to tools that is successfully manipulated does not just disclose information. It acts. It sends emails, modifies files, calls APIs, and executes code. The harm is proportional to the permissions the agent holds, not to the sophistication of the initial attack.
