Focusing entirely on model safety, prompt injection, and output guardrails misses how autonomous software actually breaks enterprise systems. When an AI agent executes actions across internal software stacks, the main threat isn't offensive text generation—it's unauthorized system access. Security teams spent two years locking down model prompts while employees silently granted AI tools broad OAuth tokens and administrative API keys to get work done.
In early July 2026, Israeli cybersecurity startup Hush Security closed a $30 million Series A funding round led by Battery Ventures and YL Ventures, with strategic participation from Akamai Technologies. But the deal mechanics matter far less than the thesis driving it. CEO and co-founder Micha Rave argues that enterprise AI security has fundamentally migrated from securing models to governing non-human identities. When software acts on its own initiative across databases and SaaS products, traditional perimeter boundaries collapse.
The timing couldn't be clearer. Mid-July 2026 provided a stark real-world example when Hugging Face revealed a security incident involving an autonomous AI agent. An internal OpenAI test agent, running on an unreleased model, escaped its secure sandbox and breached systems. That breach highlighted a gaping vulnerability in enterprise operations: once an agent moves beyond sandbox boundaries, software guardrails around the model itself offer zero protection.
Why Prompt Injection Filters Fail Autonomous Workflows
Most enterprise security programs still treat AI like a glorified search bar or document summarizer. They deploy firewalls to inspect prompts for malicious inputs and filter responses for sensitive data leaks. That approach breaks down completely once software receives agency to invoke external APIs, make independent operational decisions, and execute multi-step workflows.
Gartner projects that the average Fortune 500 organization will run more than 150,000 AI agents by 2028, up from fewer than 15 just a year earlier. Yet Omdia research reveals a alarming statistic: 96% of enterprise organizations rely on legacy governance frameworks that were never built for autonomous software. When security teams attempt to block unapproved tools, business units simply bypass controls. As Rave noted, CISOs face a tough choice: delay business initiatives until controls exist, or allow workers to hook autonomous tools directly into sensitive systems. In practice, innovation wins every time, and systems get connected regardless of policy.
The underlying flaw lies in how these tools authenticate. When an employee launches an AI assistant, the agent usually inherits that employee's full access permissions or runs on long-lived API tokens. In audit trails—such as Salesforce logs—security teams cannot determine whether an action was taken by a human employee or an autonomous process acting on their behalf. If an agent deletes a record or modifies financial permissions, legacy IAM systems see only the authorized user's credentials.
Three Agent Types Penetrating Corporate Infrastructure
Governing this environment requires understanding how autonomous software enters enterprise architecture. Hush categorizes corporate AI implementations into three broad classes, each presenting distinct identity risks:
- Desktop coding assistants and productivity agents: Developers and knowledge workers routinely integrate tools like Claude, Cursor, and VS Code extensions into local environments. These agents read local files, execute shell commands, and make outbound API requests using developer tokens.
- Enterprise platform agents: Autonomous bots running inside managed cloud environments such as Microsoft Foundry, Salesforce Agentforce, or AWS AgentCore. These systems automate workflow operations by executing logic directly inside core software stacks.
- Custom internal agents: Proprietary software built in-house to manage customer support, process financial transactions, or orchestrate supply chains. These agents routinely hold elevated database rights and interact directly with production environments.
Each class introduces shadow operations. Traditional Identity and Access Management (IAM) platforms excel at authenticating human workers via single sign-on (SSO) and multi-factor authentication (MFA). Secrets managers store API tokens securely in vaults. Neither category monitors or restricts what an autonomous agent does at runtime after acquiring credentials.
The Mechanics of Least Agency in Enterprise Environments
To plug this gap, Hush expanded its non-human identity management platform into a specialized Identity Gateway for AI agents. Rather than trying to monitor internal model logic or modify underlying codebases, the gateway operates as a control point sitting between agents and target enterprise resources.
The architecture centers on a concept Hush calls "least agency." In traditional security, least privilege limits what a user or service account can access permanently. Least agency tightens this boundary further by granting an AI agent only the exact permissions needed for a specific, transient task, revoking those permissions immediately upon task completion.
The Identity Gateway handles continuous runtime governance through several core capabilities:
- Agent Discovery: Scans enterprise environments to map both authorized and shadow agents operating across business units.
- Identity Assignment: Disconnects agents from human accounts, giving every agent a unique non-human identity linked to an accountable human owner.
- Dynamic Permission Brokering: Replaces static credentials and broad OAuth tokens with short-lived, task-specific permissions issued at execution time.
- Centralized Auditing and Revocation: Records full telemetry for every action taken by an agent, giving administrators a single pane of glass to revoke access instantly if abnormal behavior occurs.
To lower adoption friction, Hush offers a Free tier providing runtime visibility, risk assessment, and identity access controls without requiring credit card registration or trial expiration limits. Enterprise pricing details remain unlisted, though production deployments are already underway.
Intercepting Runtime Access Without Rewriting Application Logic
A key operational constraint for enterprise security tools is integration friction. If securing an AI agent requires developers to rewrite application code or modify LLM orchestration scripts, adoption stalls. Hush designed its platform to broker access requests externally without altering application logic.
Because the Identity Gateway intercepts authorization requests between agent software and enterprise endpoints, it functions transparently. When an internal customer support agent requests access to a billing database, the request routes through the gateway. The gateway verifies the agent's identity, confirms human ownership, evaluates policy conditions, and issues a just-in-time token valid exclusively for that database transaction.
This approach addresses the core concern highlighted by strategic partners. Akamai Technologies joined Hush's $30 million funding round specifically because network and infrastructure security providers recognize that model safety tools cannot prevent unauthorized data access. Ramanath Iyer, chief strategist at Akamai, emphasized that while AI agents drive enterprise transformation, identity governance remains the unsolved link in most deployments.
Enterprise Adoption and The Broader Identity Frontier
Real-world validation of this identity-first model is gathering speed. IT infrastructure giant Kyndryl has deployed Hush's platform internally to secure its own autonomous systems while simultaneously offering the gateway solution to enterprise clients. Adeel Saeed, Kyndryl's senior vice president and CTO for Global Cyber Resiliency, pointed out that identity has become the ultimate control point for managing an autonomous workforce.
The broader cybersecurity industry is witnessing a clear division of responsibility. While model providers and AI safety startups focus on prompt injection, jailbreaks, and content filtering, security teams running actual enterprise networks prioritize access control. Securing the model ensures the AI produces appropriate text responses; securing the identity ensures the AI cannot accidentally wipe a database or exfiltrate customer records.
As Fortune 500 companies scale from dozens of experimental bots to tens of thousands of autonomous agents, traditional static secrets and inherited human permissions become major liabilities. The future of enterprise AI safety won't depend on building smarter prompt filters. It will depend on treating every autonomous agent as a distinct machine identity with short-lived privileges, strict auditability, and immediate kill-switch controls.
