ModelTrust Technologies
mTrust is a managed AI firewall that protects MCP (Model Context Protocol) connections between AI agents and enterprise tools. It intercepts every request, calculates real-time behavioral trust scores, and scans for prompt injection attacks in both directions. The platform deploys in minutes and includes a global threat feed of known malicious agents and attack patterns.
- Cybersecurity
- Artificial Intelligence
- Software Only
- AI Agents
- Developer Tools
Funding
Founders
Founder details are not available yet.
Product
Problem
MCP (Model Context Protocol) is becoming the standard way AI agents interact with software, yet it has no built-in security layer. When an agent connects to an MCP server, there is no identity check, trust score, behavioral analysis, or audit trail — the server simply executes whatever the agent requests. A compromised agent can exfiltrate data, execute unauthorized commands, and pivot through connected systems with 80-90% autonomy and zero human oversight.
Solution
mTrust is a managed AI firewall that sits between AI agents and tools, intercepting every request. It identifies the agent, calculates a real-time trust score based on behavioral patterns, enforces security policies, and learns what "normal" looks like for each environment. The platform protects in both directions: it shields MCP servers from untrusted agents and safeguards agents from malicious MCP servers. By combining trust scoring, injection scanning, and a global threat feed, mTrust catches attacks that rule-based systems miss. The solution deploys in minutes and provides continuous monitoring and adaptive policy enforcement as agent behavior evolves.
Target Audience
Primary customers are engineering and security teams at organizations deploying MCP servers or building AI agents, including enterprises using ChatGPT, Claude, or other AI assistants to access internal tools and data.
Features
- Real-time trust scoring for AI agents based on behavioral analysis and historical patterns, with sub-2ms request interception
- AI-powered prompt injection scanning (via Bedrock) that detects and blocks malicious instructions, including hidden instructions
- Global threat feed aggregating 15+ real-world attack patterns with auto-deny rules and cross-customer agent reputation tracking
- Bidirectional protection covering both MCP server-to-agent and agent-to-server traffic
- Behavioral learning engine using parameter embeddings (Bedrock Nova) and Markov chain analysis to detect sequence attacks and anomalous tool arguments
- Tool description rug pull detection via SHA-256 hash verification on every call
- Browser-side agent verification through a 1.4KB SDK (npm, CDN, WordPress plugin)
- OAuth 2.1 with trust-gated scoping where scope is determined by behavioral trust score
- Slack interactive alerts for acknowledging or blocking agents, plus SIEM integration (CEF format) for SentinelOne, Splunk, Elastic, and CrowdStrike
- Audit trail and logging for every agent request, enabling post-incident investigation and compliance reporting