Security for the Agent Era
Trusteed is built from the ground up to operate safely in multi-agent environments. This page documents our threat model, defenses, and responsible disclosure process.
Threat Model
Six threat categories we actively defend against in the MCP ecosystem.
Prompt Injection
Malicious instructions embedded in product content, descriptions, or third-party data that attempt to hijack agent behavior.
INTEGRITY NOTICE in llms.txt; all tool inputs validated via Zod schemas; no raw HTML in agent-facing responses.
Tool Poisoning
An MCP tool returns a crafted payload designed to alter the agent's subsequent actions or exfiltrate context.
Signed tool responses; strict output schema validation; tool output never interpolated into system prompts.
Checkout Abuse
An agent—or attacker posing as one—creates carts and checkouts at high volume to exhaust inventory or trigger fraud.
Rate limiting enforced: create_cart 20 req/min, complete_checkout 10 req/min per agent identity.
Fraudulent Agents
A rogue agent presents itself as a legitimate buyer to bypass fraud controls or access restricted operations.
OAuth 2.0 M2M authentication required; KYAI engine applies merchant-defined rules per agent identity.
Merchant Impersonation
A fake platform mimics a legitimate merchant to intercept agent transactions or capture payment credentials.
Trust score >= 0.70 required for checkout-eligible merchants; QTSP timestamps on all signed trust records.
Data Exfiltration
An agent attempts to extract customer PII, order history, or payment data beyond its authorized scope.
Scoped API keys enforced per operation; no PII included in agent-facing tool responses; agent_events log all data access.
API Key Scopes
All agent access is restricted to declared scopes. Requesting data outside a key's scope returns a 403.
read:productsSearch and read product catalogread:merchantView merchant profile and trust scorewrite:cartCreate and modify cartwrite:checkoutInitiate checkout flowadmin:merchantStore management — merchants onlyAudit & Confirmation Model
security-logger.ts records every security event: auth failures, rate limit hits, scope violations, and anomalous patterns.
Merchants configure human-confirmation rules via the KYAI engine — e.g. require approval for orders above a threshold.
Agent event logs are available in the merchant dashboard (agent_events table), filterable by agent identity and action type.
Responsible Disclosure
We take security reports seriously and respond quickly to critical vulnerabilities.
- Security email
security@trusteed.xyz - Response SLA
< 72 hours for critical vulnerabilities - Trust appeal
POST /api/v1/trust/appeal