Your Devs Run AI Agents.
Your Production Systems Run AI Agents.
Nobody's Watching.

Runtime security & compliance governance for the AI agent era

Development

Cursor · Claude Code · Windsurf · Copilot · GitHub Copilot Workspace

Executes shell commands with dev's full permissions

Reads ~/.ssh, ~/.aws, env vars, entire filesystem

Writes to any file, installs packages, runs scripts

Production

OpenClaw · LangChain · OpenAI Agents SDK · CrewAI · AutoGen · Anthropic Agent SDK

Queries and modifies production databases

Calls external APIs with real credentials

Processes financial transactions autonomously

Accesses customer PII without audit trail

The attack surface

This isn't theoretical. It's already happening.

$250K Lost in Seconds

Lobstar Wilde: AI agent sent 52M SOL instead of 4. No numeric bounds check. No runtime validation.

No numeric bounds

SSH Key Exfiltration

Prompt injection via PR comment caused coding agent to exfiltrate developer SSH keys to attacker server.

Prompt injection → credential theft

MCP Supply Chain Attack

Poisoned MCP tool descriptions injected hidden instructions, hijacking agent behavior through tool metadata.

Tool description poisoning

Bulk Data Exfiltration

Production agent entered agentic loop, exfiltrating customer data record-by-record without triggering any alerts.

Agentic loop → exfil

Production DB Wiped

Coding agent misinterpreted “clean up the schema” and ran DROP TABLE on production database. No rollback, no confirmation gate, 1.2M rows gone.

Destructive query → no confirmation

Agent Identity Spoofing

Attacker embedded hidden instructions in a support ticket. Customer-facing agent escalated its own privileges and issued a $47K refund autonomously.

Indirect prompt injection → privilege escalation

EU AI Act enforcement: 2026 deadline. Non-compliance fines up to €30M or 6% of global turnover. Every company deploying AI agents in the EU needs governance — now.

Product suite

Three products. One governance platform.

AgentShield

Runtime Security Gateway

• 6-layer analyzer pipeline: regex → structural → semantic → dataflow → stateful → guardian
• BLOCK / AUDIT / ALLOW decisions in real-time
• IDE hooks: Claude Code, Cursor, Windsurf, OpenClaw
• MCP mediation: tool description poisoning, content scanning, value limits

100% accuracy on 147 test cases (123 shell + 24 MCP)

Comply

AI Risk Compliance Platform

• Scans codebases for AI usage (Semgrep-powered, 37 rules, 6 languages)
• Maps to 5 frameworks: OWASP LLM Top 10, EU AI Act, SOC 2, ISO 27001, NIST AI RMF
• Auto-generates AgentShield policy packs
• 13-section audit-ready compliance reports
• React dashboard for visualization

37 Semgrep rules · 5 frameworks · 6 languages

Rules Hub

Security Rule Management

• Browse and edit static rules, runtime packs, taxonomy
• Compliance coverage matrix across all frameworks
• Search across all rule types
• Version-controlled rule lifecycle
• Community-contributed rule database

32 taxonomy entries · 7 threat kingdoms

The governance loop

Scan. Assess. Enforce. Audit. Repeat.

Scan Code

Semgrep rules find AI usage

→

Assess Risk

Map to compliance frameworks

→

Generate Policy

Auto-create enforcement rules

→

Enforce

AgentShield runtime gateway

→

Audit Log

Every action recorded

→

Report

Compliance-ready output

→

Rescan

Continuous governance

Why this is different

SAST did this for code.
We're doing it for AI agents.

Key differentiator: not just finding problems, but generating and enforcing the fix automatically.

Defensibility

The Rules Are the Product

Every scan enriches the rule database. Compliance intelligence compounds.

Compounding Intelligence

37 Semgrep rules, 32 taxonomy entries, 5 compliance frameworks — all open source, all growing with every deployment.

Network Effects

A startup plugging in on day one inherits the compliance intelligence of every company before it.

Hard-Won Domain Knowledge

Rules are domain expertise encoded as code. This protects us from new competitors — rules are hard-won knowledge, not just software.

By the numbers

Built and battle-tested.

OWASP LLM Top 10 2025 fully aligned

Market opportunity

$308M today. $3.6B by 2033.

Target segments

Team

36+ combined years in security & engineering leadership

Gary

Co-founder & CTO

• 18+ years Application Security (AppSec)
• CISSP, CSSLP certified
• Built SAST detection rules for enterprise software — now applying the same model to AI agent runtime security
• Led AI/LLM attack surface research: prompt injection, jailbreaks, data poisoning
• Cloud security: ISO 27001, PCI-DSS, SOC compliance

Anshuman Biswas, PhD

Co-founder & CEO

• 18+ years Engineering Leadership
• VP Engineering at Elastio — ransomware detection, AI-powered anomaly detection, 100M+ files/day
• 6+ years Turbonomic/IBM — orchestration, multi-team management
• PhD Systems & Computer Engineering, Carleton University
• 6 published papers on cloud auto-scaling and ML

Let's Build the Governance Layer
for AI Agents

Open source. Enterprise ready. Start protecting your AI agents today.

[email protected] [email protected]