# AI runtime security glossary

URL: https://airuntimesecuritytools.com/glossary/
Updated: 2026-11-16

Twenty-eight terms define the vocabulary of AI runtime security, from fail-open and fail-closed to decision latency, output filtering, tool-call authorization, reverse proxy, sidecar deployment and false positive budget. Each entry gives a definition and one sourced example read on 16 Nov 2026.

## 28 terms

- [Anomaly Detection](/glossary/anomaly-detection/): Anomaly detection flags activity that departs from an expected pattern without matching a known bad signature.
- [Behavioral Baseline](/glossary/behavioral-baseline/): A behavioral baseline is a learned picture of what an agent, or the person who runs it, normally does, such as which files it touches and which tools it calls.
- [Block vs Detect](/glossary/block-vs-detect/): Block versus detect is the choice between stopping an action before it completes and only raising an alert.
- [Canary Token](/glossary/canary-token/): A canary token is a fake credential, file or setting planted where nobody should look, which raises an alarm the moment it is touched.
- [Decision Latency](/glossary/decision-latency/): Decision latency is the time an inline security check takes to return allow, block or redact for one prompt, reply or action.
- [eBPF](/glossary/ebpf/): eBPF is a Linux kernel technology that runs small sandboxed programs inside the operating system to observe or control activity without changing kernel code.
- [Fail Closed](/glossary/fail-closed/): Fail closed describes a control that blocks traffic or actions when the security check is unavailable, so a failure never opens a gap.
- [Fail Open](/glossary/fail-open/): Fail open describes a control that lets traffic or actions through when the security check is unavailable, slow or times out.
- [False Positive Budget](/glossary/false-positive-budget/): A false positive budget is the number of wrongly blocked or flagged safe actions a team agrees to tolerate in exchange for catching real attacks.
- [Inline Enforcement](/glossary/inline-enforcement/): Inline enforcement means the security check sits in the path of the request, so the action waits for a verdict and can be blocked.
- [Input Filtering](/glossary/input-filtering/): Input filtering inspects what is sent to a model, including user messages and retrieved documents, and can block or change it before the model sees it.
- [Jailbreak](/glossary/jailbreak/): A jailbreak is an input crafted to make a model ignore its safety rules and produce content or take actions it was set up to refuse.
- [Kernel-Level Monitoring](/glossary/kernel-level-monitoring/): Kernel-level monitoring watches the operating system's own events, such as system calls, process starts and file opens, from the most privileged part of the machine.
- [Model Denial of Service](/glossary/model-denial-of-service/): Model denial of service is an attack or accident that overloads an AI model with requests or oversized inputs until the service slows, fails or runs up a bill.
- [Output Filtering](/glossary/output-filtering/): Output filtering inspects what a model returns and can block, mask or rewrite it before a user or another system receives it.
- [PII Redaction](/glossary/pii-redaction/): PII redaction replaces personal data in a prompt or reply with a placeholder or mask before it leaves the organization or reaches a user.
- [Rate Limiting](/glossary/rate-limiting/): Rate limiting caps how many requests may reach a service in a time window and rejects the rest.
- [Reverse Proxy](/glossary/reverse-proxy/): A reverse proxy for AI traffic is a gateway that sits in front of model providers and forwards requests after inspecting them.
- [Runtime Protection](/glossary/runtime-protection/): Runtime protection is a security control that watches an AI system while it is running and can stop a harmful prompt, reply or action before it completes.
- [SDK Integration](/glossary/sdk-integration/): An SDK integration adds the security check inside the application's own code through a vendor library that calls the service before and after model or tool calls.
- [Sensitive Information Leakage](/glossary/sensitive-information-leakage/): Sensitive information leakage is the exposure of personal data, credentials, business secrets or model internals through an AI system's output.
- [Sidecar Deployment](/glossary/sidecar-deployment/): A sidecar deployment runs the security check as a small companion process beside each application or agent, on the same host or pod.
- [SIEM Export](/glossary/siem-export/): SIEM export is the sending of a runtime tool's alerts and decision records to a security information and event management system.
- [SOAR Playbook](/glossary/soar-playbook/): A SOAR playbook is an automated set of response steps that runs when an alert or incident matches a rule.
- [System Prompt Leakage](/glossary/system-prompt-leakage/): System prompt leakage is the exposure of the hidden instructions that steer a model, which can reveal rules, roles and any secrets placed in them.
- [Telemetry](/glossary/telemetry/): Telemetry is the stream of records a runtime tool emits about what it saw and decided: prompts, tool calls, verdicts, timing and the policy that applied.
- [Token Budget](/glossary/token-budget/): A token budget is a cap on how many model tokens a user, team, app or agent may consume in a period.
- [Tool Call Authorization](/glossary/tool-call-authorization/): Tool call authorization is a decision, made before an AI agent runs a tool, on whether this agent may call this tool with these arguments.

## Related pages

- [Guides](/guides/)
- [FAQ](/faq/)
- [The ranking](/ranking/)
- [Runtime Action Type Matrix](/runtime-action-type-matrix/)
- [Enforcement Point Map](/enforcement-point-map/)
