AI security glossary.
The vocabulary of AI security, in plain language: prompt injection, LLM jailbreaks, firewalls, guardrails, data loss prevention, agent security, MCP, model supply chain and the evidence that ties a decision to a record. Each entry links to the part of the product that handles it.
Prompt injection
An attack in which text a model is meant to read as data — a document, a web page, a tool result — contains instructions the model follows as if they came from its operator. Direct injection arrives in the user prompt; indirect injection arrives through retrieved content. A prompt injection defence screens both before the model is called.
The engines that detect it →LLM jailbreak
A jailbreak persuades a model to ignore its own safety policy, usually through roleplay, encoding, or a long multi-turn story. It is not the same as prompt injection: injection hijacks the task, a jailbreak removes the guardrail. Detection has to read intent and structure across the conversation, not match one keyword.
Jailbreak and injection protection →LLM firewall
An inline control that inspects a prompt before it reaches a model and the response before it reaches the user, then allows, flags or blocks the call. A firewall produces a verdict and evidence for every request, rather than a periodic report.
How the gateway fails closed →AI guardrails
Policy constraints applied to model input, output and agent actions. Guardrails can prevent an action, redact data, or require human approval. Detection without enforcement is monitoring; a guardrail changes what the system is allowed to do.
Policy packs and two-approver promotion →AI data loss prevention (DLP)
Inspecting AI traffic for personal data, credentials, source code and regulated information, then redacting or blocking it. DLP has to cover both directions: what the user sends and what the model returns.
Data loss prevention engine →PII redaction
Replacing personal identifiers with placeholders before data crosses a boundary, so the model can still complete the task without holding the identity. Redaction should be reversible only by the system that owns the data.
What the engine stores →AI agent security
Governing autonomous agents: which tools they may call, what they may spend, which goals they may pursue, and when a human must approve. Agents fail differently from chatbots because their actions have side effects outside the conversation.
Agent action policy and drift →MCP security (Model Context Protocol)
The Model Context Protocol connects models to tools and data sources. Securing it means auditing tool descriptions for poisoning, constraining scopes, and detecting confused-deputy calls where a trusted server is used to reach something it should not.
MCP and tool governance →AI model supply chain security
Verifying where a model came from and that it is the model you think it is: provenance, hashes, formats and known-bad artefacts. A tampered model can behave correctly in testing and differently in production.
Model supply chain engine →AI red teaming
Adversarial testing of an AI system with the prompts, tool calls and multi-turn strategies an attacker would use. The useful output is evidence of which attacks succeed and which controls stopped them, not a single score.
Run an adversarial scan →Denial of wallet
An attack that makes an AI system consume far more tokens or API calls than intended — recursive prompts, long loops, or abusive clients — until the bill or a rate limit stops it. Controls need per-session ceilings, not just a global quota.
Wallet guard →AI audit trail
A durable, ordered record of what the system decided and why: input hash, verdict, policy, action and evidence. A tamper-evident trail chains each record to the one before it, so a later edit is detectable rather than invisible.
The evidence model →Proof of intent
A signed statement that a specific human authorized a specific action within a defined scope, at a known time. It turns “the agent did it” into a verifiable claim about who approved it and what they approved.
Verifiable proof of intent →AI escrow
Holding an agent's funds or permissions until an independent check confirms the work was done. Escrow limits the damage a compromised or mistaken agent can do with money or authority.
Agent escrow pricing →AI content moderation
Screening model inputs and outputs for spam, sexual content, minor safety, profanity, violent threats, self-harm and hate, with a category and a confidence for each. The engine supplies the evidence; the policy decision stays with you.
Content moderation engine →OWASP Top 10 for LLM Applications
A community risk taxonomy for LLM applications that includes prompt injection, sensitive information disclosure, supply chain risks, excessive agency and unbounded consumption. It is a map of risks, not a certification; Intellirity's engines address several of its categories and produce evidence for each verdict.
See the engine coverage →