Runtime guardrails for LLM apps and AI agents
Scan what goes into and comes out of your model, and what your agent does with tools, against a policy you control.
Challenge and solution
Challenge
Prompts carry injected instructions, responses leak personal data and secrets, and tool results can steer an agent. Filtering only the user's prompt leaves most of that path open.
Solution
AI Protect scans six stages of an LLM call against project policy: prompt, response, tool call, tool result, retrieval and document. You reach it through a guard API or an OpenAI-compatible gateway, with the Python SDK or REST.
Key capabilities
Six stages, one policy
Prompt, response, tool call, tool result, retrieval and document, so an injection hidden in a retrieved page or a tool result is scanned too.
16 finding categories
Injection, unsafe link, code, gibberish, PII, secret, harmful content, child safety, language policy, custom rule, denied topic, competitor, off-topic, plain-English rule, system prompt leak and quarantine.
Classifier plus guard model
A classifier handles the clear cases and a guard model reviews the uncertain ones. Three sensitivity levels, and 11 languages including Arabic.
PII and secrets
PII detection uses checksum validators, and replace-and-restore swaps values out before the model and puts them back after. 214 secret rules look for keys and tokens.
Repeat-offender quarantine
Callers that repeatedly trigger findings can be quarantined.
Templates and custom detectors
Five policy templates (support, coding, internal knowledge, research, finance). Add custom detectors of six types: regex, heuristic, PII, classifier, embedding and package registry.
Versions, shadow test and rollback
Each policy change is a new version. Shadow test it before it enforces, and roll back if needed.
Content-free events and zero retention
Events record what was found, not the text. Zero-retention mode is available for supported detection events.
What you get
Injection and leakage caught in line
Findings are returned from the guard call, so your app can block, redact or log before the next step.
Agent steps covered, not just chat
Tool calls, tool results and retrieved content are scanned, which is where agent attacks usually enter.
Policy changes without guesswork
Shadow test shows what a new version would catch before it enforces.
Less sensitive data in your logs
Content-free events and zero-retention mode keep prompt text out of stored events.
Frequently asked questions
How do I connect my app?
Call the guard API from the Python SDK or REST, or point your OpenAI-compatible client at the gateway. You choose which of the six stages to scan.
Does Aynigma store my prompts?
Events are content-free, meaning they record the finding and not the text. Zero-retention mode is available for supported detection events. Check the current data-handling terms with us before sending regulated data.
How do I change a policy safely?
Policies are versioned. Run a new version in shadow test before it enforces, then roll back if it blocks too much.
Put guardrails in front of your LLM app.
Tell us which stages you need to scan and how your app calls the model. We will confirm the integration route and data-handling terms.