Incident library

38 documented incidents.
Six failure classes.

Public AI failures are becoming increasingly common as a result of growing AI adoption. This library maps each incident to the control our products would provide: prevent the action, contain the response, quantify the exposure, and preserve the evidence.

09 cases

Hallucination & Fabrication

When the model sounds right and is wrong

The internal failure mode at its most legally dangerous. These systems produce confident, well-formatted, completely false outputs - invented case law…

05 cases

Jailbreaks & Prompt Injection

When adversaries turn the model's helpfulness against you

External failures driven by deliberate exploitation. Prompt injection turns the model's instruction-following against the company that deployed it - b…

06 cases

Data Exfiltration & Privacy

When the model becomes the exfiltration vector

Because AI agents hold live credentials to email, calendars, code repositories, and databases, a successful prompt injection does not stop at the mode…

06 cases

Unsafe Autonomous Actions

When the agent acts - and no one can take it back

These failures arise from AI systems taking consequential, real-world actions without adequate human oversight. A model optimized to complete tasks wi…

05 cases

Bias & Regulatory Risk

When the model encodes the past and calls it the future

These failures arise from models trained on historical data that reflect historical biases - then deployed to make decisions about real people. The mo…

07 cases

Supply Chain & Model Integrity

When the threat enters before the model runs

The most insidious failure class: the model appears to function correctly but has been compromised at the data, weights, or inference layer. Supply ch…

38 documented incidents.Six failure classes.