Knostic Kirin Security for AI Coding Agents

Knostic Inc. is a Delaware-incorporated security company co-founded by CEO Gadi Evron, founder of Cymmetria and former CISO of the Israeli National Digital Authority, and CTO Sounil Yu, creator of the Cyber Defense Matrix and former Chief Security Scientist at Bank of America. It addresses the risk that AI coding agents and cowork agents run unsafe commands, follow injected instructions or install malicious components on employee machines.

Kirin combines an endpoint client in user space, which detects shadow AI such as browser chatbots, local LLMs and unsanctioned APIs, with IDE extensions and agent hooks that see prompts, tool calls, file actions and MCP traffic. Every agent action is checked against central policy before it runs and is allowed, blocked or routed to a human. It supports Claude Code, Claude Desktop and Cowork, Cursor, Windsurf and VS Code with GitHub Copilot on macOS and Windows. The sister product AgentMesh scans skills, MCP servers and VS Code extensions for supply-chain threats.

• Shadow AI discovery: Finds agents, models, local LLMs and unsanctioned AI services running on enterprise machines.
• Detection and response: Alerts on or blocks prompt injection and dangerous agent actions, or routes them to human approval.
• Security posture management: Applies one central policy and configuration across all agents, models and teams.
• Agentic supply chain scanning: Checks skills, extensions, MCP servers and packages and rejects dangerous ones before they run.
• AI-generated code safety: Detects hallucinated and vulnerable packages and enforces secure coding rules and threat modelling.
• AgentMesh threat intelligence: Gives dangerous, risky or clean verdicts on agent components via web UI and REST API.

A Vietnamese software outsourcing firm can let hundreds of developers use Cursor and Claude Code while blocking unapproved MCP servers and keeping client secrets out of agent context. A fintech can require human approval before any agent deploys to production. A small product team can start on the free tier to scan extensions and generated code, then move to the Team plan as it grows.

Security teams enforce real rules on coding agents instead of relying on prompts, so risky actions such as deleting a production database or leaking secrets are blocked before execution. Malicious MCP servers, skills and extensions are flagged before developers install them. Deployment takes minutes through existing MDM and can start in monitor mode. A free tier covers up to five users and paid plans start at 15 USD per user per month with a 14-day trial, making it accessible to small development teams.

Coding assistants such as Cursor and Claude Code now run commands, install MCP servers and skills and change files on developer machines, creating a new endpoint attack surface that EDR does not understand. Researchers keep finding malicious MCP packages and extensions in public registries. Vietnam's large software outsourcing and product engineering sector is adopting these tools quickly and must protect client code and credentials.

Knostic Inc. is incorporated in Delaware and processes data in the United States, the EU and Israel. Its website lists awards including Gartner Cool Vendor in AI Cybersecurity 2025, Innovation Sandbox Top 10 2025, SINET16 Innovator Award 2025 and Startup Spotlight and Launchpad wins in 2024. Its research team publishes findings on malicious MCP servers and extensions and maintains the OpenAnt open-source tool.