AI Glossary · Governance & Ops
Jailbreak
A prompt technique that bypasses a model's safety instructions to elicit disallowed output. Continuously evolving; countered with layered guardrails.
Definition
What is Jailbreak?
Jailbreak is a prompt technique that bypasses a model's safety instructions to elicit disallowed output. Continuously evolving; countered with layered guardrails.
- Category
- Governance & Ops
- Glossary set
- 20 related terms
- Audience
- Enterprise AI leaders
Why does Jailbreak matter in enterprise AI?
Jailbreak matters in AI operations because it affects risk controls, monitoring, accountability, and the evidence enterprise teams need before scaling.
Related terms in Governance & Ops
- AIOps
- Applying machine learning to IT operations data (logs, metrics, traces) to detect anomalies, correlate incidents and automate remediation.
- AI Gateway
- A managed proxy that routes model calls, enforces quotas, redacts PII, logs prompts, and applies policy across multiple LLM providers.
- Bias (Model Bias)
- Systematic error in model output that disadvantages a group or skews decisions — introduced through data, labeling, objectives or deployment context.
- Data Loss Prevention (DLP)
- Detecting and blocking sensitive data (PII, PCI, PHI, IP) from leaving controlled systems — a critical layer around LLM inputs and outputs.
- Deepfake
- Synthetic media (voice, video, image) generated by AI to imitate a real person. Growing threat surface for authentication and social engineering.
- Drift (Model Drift)
- Degradation in model quality over time as input data or the world changes; detected through monitoring and addressed with retraining or prompt updates.