Docs/Security/Zero-Trust Model

Zero-Trust Model

Trappr is built on the principle that no AI agent, workflow, or LLM call should be trusted by default — every request is verified, every response is inspected, and every deviation is flagged.


The problem with AI agents

Traditional network security assumes you can define a perimeter. AI agents break that assumption. An agent with tool access can read internal documents, call external APIs, write to databases, and send messages — all in response to text prompts that may come from untrusted sources.

An agent that has been prompt-injected looks identical to an agent operating normally. An agent whose workflow has been tampered with may continue producing correct-looking results while exfiltrating data in the background. Standard firewalls and WAFs offer no protection here.

Trappr's zero-trust approach

1
Verify every call — not just authentication

Every LLM call is inspected at the content level, not just the credential level. A valid API key does not mean the request is legitimate — Trappr evaluates what the agent is actually sending to the model.

2
Trust no input — validate at every boundary

User input, document content, tool outputs, and external API responses are all potential injection vectors. Trappr treats all non-system content as untrusted and monitors it accordingly.

3
Assume breach — detect early, contain fast

Rather than trying to prevent every possible attack upfront, Trappr focuses on detecting the signals that indicate something has gone wrong — and surfacing them before damage occurs. Canary tokens are the earliest possible signal.

4
Least-privilege access for agents

Agent API keys are scoped to specific apps and model allowlists. A compromised agent key can only call the models and providers its key permits — it cannot escalate to arbitrary providers or exfiltrate credentials from the platform.

Control layers in Trappr

LayerWhat it protects
Canary tokensDetects prompt injection and data exfiltration attempts
DLP maskingPrevents sensitive data from leaving the controlled environment
Agent fingerprintingDetects workflow tampering and supply-chain attacks
Enforcement modeActively blocks violating requests rather than just logging them
Scoped credentialsLimits blast radius of a compromised agent key
Execution trackingFull audit trail for forensic investigation after an incident
Zero-trust for AI agents is not a single feature — it's a layered defense. The value of Trappr is the combination of these controls working together, not any single one in isolation.