Adversa AI Agentic Red Teaming and Coding Agent Security

Adversa AI Ltd. is a Tel Aviv-based AI security company co-founded by CEO Daniel Rubinstein and CTO Alex Polyakov, who has about twenty years of offensive security experience and co-leads the Agentic AI Security initiative at the Coalition for Secure AI. It holds US patent 11,275,841 B2 on AI protection and maintains a knowledge base of more than 4,000 adversarial AI papers. It works with Fortune 500 companies and governments in the US, EU and Asia.

Its continuous red teaming platform tests agentic AI systems, GenAI applications, models and MCP servers for goal hijacking, prompt injection, tool poisoning, cross-tenant leakage, zero-click RAG exploits, command injection and OAuth bypass, delivered as threat modeling, continuous automated assessment and advanced multi-stage red teaming. A newer Coding Agent Security Platform combines a gateway that sees LLM and MCP traffic with an endpoint sensor on developer machines to block dangerous action chains by Claude Code, Copilot, Cursor and Codex.

• Continuous agentic red teaming: tests autonomous agents, GenAI apps and models for prompt injection, tool leakage and goal hijacking.
• MCP security testing: covers tool poisoning, rug pulls, tool shadowing, cross-tenant leakage and OAuth bypass.
• Threat modeling: architecture analysis of agent and MCP deployments.
• Multi-stage attack chains: advanced red teaming with conditional payloads designed to evade detection.
• Coding agent runtime control: gateway plus endpoint sensor that scores and blocks risky action chains.
• AIRQ research: risk ratings of 100 enterprise AI agents to inform procurement.

Vietnamese software outsourcing companies using AI coding agents can enforce runtime controls to protect client code and secrets. Banks and fintechs connecting agents to internal systems via MCP can red team those integrations before launch. Telecom operators can continuously test customer-facing GenAI services.

Enterprises find exploitable weaknesses in agents and MCP integrations continuously rather than through one-off audits. Testing runs on the customer's infrastructure and the MCP offering states no data access is required. The coding agent platform lets engineering teams adopt AI coding tools while warning, denying or escalating risky actions such as deleting production data or leaking secrets. SOC 2 Type I and II audits and ISO 27001 certification support vendor due diligence.

Agentic AI and MCP have created new attack classes such as tool poisoning and zero-click exploits, and AI coding assistants now act directly on source code and production systems. Software firms and enterprises across Asia that adopt these tools need continuous testing rather than one-time model evaluations.

Adversa AI Ltd. lists its address at Rothschild Boulevard 45, Tel Aviv. It reports SOC 2 Type I and II audits, ISO 27001 certification, over 20 industry awards including the 2025 Fortress Cybersecurity Award and a 2025 Gold Stevie, and recognition from IDC, Gartner, Wavestone and KuppingerCole. It created the open-source SecureClaw agent security framework.