# Arrakis raises $8 million for AI agent runtime security > Source: > Published: 2026-08-02 10:32:56+00:00 [Tal Baron](https://il.linkedin.com/in/tal-bar-on?ref=runtimewire), [Omer Efrat](https://www.linkedin.com/in/omerefrat/?ref=runtimewire) and [Ron Shani](https://il.linkedin.com/in/ron-shani-8a264673?ref=runtimewire) disclosed an $8 million seed round for [Arrakis Security](https://arrakis.security/?ref=runtimewire) on August 2, giving the Torq and Palantir veterans capital to build runtime controls for enterprise AI agents, [CTech reported](https://www.calcalistech.com/ctechnews/article/hyklcf2bgg?ref=runtimewire). [Hetz Ventures](https://www.hetz.vc/?ref=runtimewire) led the round. The investors also included [ElevenLabs](https://elevenlabs.io/?ref=runtimewire) CEO Mati Staniszewski, [Torq](https://torq.io/?ref=runtimewire) CEO Ofer Smadari, [Pentera](https://pentera.io/?ref=runtimewire) CEO Amitai Ratzzon and several Palantir executives whose names were not listed by CTech. Arrakis employs about 20 people and plans to expand research, development and product hiring in Israel, according to the report. Baron and Efrat came out of Torq, while Shani spent eight years at Palantir. Efrat, a former Torq vice president of product, has worked in product management since 2010 after seven years as a developer, according to a [ProductX profile](https://productx.org/board/item/503-omer-efrat-vp-product-torq?ref=runtimewire). The backgrounds fit the problem Arrakis has chosen. Torq's public product work centered on security automation, while Shani worked inside a software organization known for deploying complex data systems into large institutions. Their new bet is that autonomous agents create a separate control problem once software can use credentials, call tools and change business systems without waiting for a person to approve each step. ### Security moves from identity to behavior Arrakis calls its product an "Autonomous Workforce Governance" platform. The language treats AI agents as non-human workers carrying permissions that previously belonged to employees or conventional service accounts. The [Arrakis platform](https://arrakis.security/?ref=runtimewire) is designed to discover sanctioned and shadow agents, connect each agent to an owner, inventory its permissions, establish behavioral baselines and calculate risk scores. Arrakis says it covers workflow systems such as n8n, Make, ServiceNow and Salesforce Agentforce; coding tools including Cursor, Claude Code, Devin and GitHub Copilot; and desktop assistants from the leading AI providers. Arrakis' main distinction is what happens after an agent receives access. Its controls are designed to inspect tool calls and activity across endpoints, SaaS applications and cloud infrastructure. When behavior violates policy, Arrakis says it can block an action, revoke permissions, suspend an integration or activate a kill switch. That pitch pushes beyond inventory and identity management into detection and response. The [Arrakis documentation](https://arrakis.security/docs?ref=runtimewire) describes an architecture that connects identity providers, device-management systems, developer endpoints and security information and event management systems. Its [MCP gateway](https://arrakis.security/docs/platform-concepts/mcp-gateway/?ref=runtimewire) inspects traffic between agents and external tools, applies allow-lists and data-loss controls, and checks tool-server packages for known vulnerabilities. Arrakis also offers [red-teaming capabilities](https://arrakis.security/docs/platform-concepts/red-teaming/?ref=runtimewire) aimed at multi-turn manipulation, excessive permissions and attacks that an agent cannot reliably identify inside its own reasoning process. The premise is straightforward: an autonomous system cannot serve as the only judge of whether its behavior has been manipulated. ### Hetz is placing adjacent bets Hetz's investment is also notable because the fund has backed another young Israeli startup tackling a closely related problem. On June 4, [Willow announced a $7 million seed round](https://withwillow.ai/blog/willow-7m-seed-funding?ref=runtimewire), also led by Hetz, to build governance software for enterprise AI agents. Arrakis emphasizes behavioral monitoring, threat detection and automated containment across the agent, its workflow and the surrounding SaaS graph. Enterprise security budgets rarely divide into categories as neatly as vendor diagrams do, leaving AI-agent security companies likely to encounter the same buyers and deployment questions. Hetz's willingness to fund adjacent approaches reflects how early the category remains. Security buyers have not settled on whether agent governance belongs inside identity management, endpoint security, AI gateways, cloud security or security operations. Founders are racing to establish the control point before larger platforms absorb these functions. The market already includes evidence of that consolidation. [Check Point acquired Cyata in February 2026](https://www.checkpoint.com/de/press-releases/check-point-software-reports-fourth-quarter-and-2025-full-year-results/?ref=runtimewire) for agent discovery, governance and control. [CodeIntegrity raised $5 million in May](https://www.codeintegrity.ai/blog/codeintegrity-seed-round?ref=runtimewire) to govern AI agents running in production. SentinelOne agreed to [acquire Prompt Security in August 2025](https://www.sentinelone.com/press/sentinelone-to-acquire-prompt-security-to-advance-genai-security/?ref=runtimewire) as part of its generative AI and agent-security strategy. ### The seed round buys time to prove the control layer Arrakis is entering the market before enterprises have proved how much autonomous software they will run in production. [Gartner predicted in June 2025](https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027?ref=runtimewire) that more than 40% of agentic AI projects would be canceled by the end of 2027 because of rising costs, unclear business value or inadequate risk controls. That warning cuts both ways for Arrakis. Failed deployments could shrink the number of agents requiring protection. Stronger governance may also be what moves useful projects beyond prototypes and internal experiments. Baron, Efrat and Shani are betting that enterprises will want a dedicated security layer between an agent's permission to act and the action itself. Their experience gives Arrakis credibility with security teams and complex enterprise deployments. The harder task starts after the seed announcement: proving that runtime oversight can catch meaningful agent behavior without obstructing the work those agents were deployed to perform.