Radware unveils Agentic AI Protection for autonomous agents

Radware has launched an Agentic AI Protection Solution, a new security product aimed at organisations deploying autonomous AI agents in business systems and workflows.

The company positioned the release as an expansion of its platform into AI security. The product targets risks that arise when AI agents interact with tools, data stores, email systems and other enterprise applications without direct human oversight.

AI spending continues to draw attention from security vendors. Gartner Research forecast worldwide AI spending will reach $2.5 trillion in 2026, including $51.3 billion for AI security solutions, according to Radware.

"We have seen the AI security wave coming for many months, and knew we needed to support our customers in providing the security surrounding it," said Roy Zisapel, President and CEO, Radware.

"Our customers understand the dual nature of AI - they want the business benefits while recognizing the growing threat landscape. With decades of experience applying advanced algorithms to solve complex application security challenges, Radware is uniquely positioned to bring innovative AI security solutions to market."

Runtime focus

Radware said many current AI security approaches use static guardrails. These typically focus on governance and policy controls. The company argued that these measures do not address the behaviour of autonomous, tool-using agents during live operation.

Agentic AI Protection uses behavioural analysis outside the agent environment, Radware said. The company said this approach can identify malicious intent and misuse in real time.

The product addresses a set of agent-specific threats. Radware highlighted direct and indirect prompt injection, tool abuse, human-agent trust exploitation and unauthorised data access.

Four pillars

The product is built around four elements that Radware described as strategic pillars.

The first is discovery and visibility. Radware said it identifies AI agents in real time, including internally built agents and software-as-a-service agents. It also maps the tools and systems those agents can access.

The second is intent-based security. Radware said it uses runtime behavioural algorithms that detect and mitigate malicious or abnormal intent within agent interactions. The company said this includes multi-step behaviours and patterns that span multiple agents.

The third is integration with agent platforms and services. Radware said the product can protect custom-built agents and third-party services. It listed Microsoft 365 Copilot, Microsoft 365 Copilot Studio and AWS Bedrock among the services it supports, alongside in-house agents.

The fourth is continuous posture management. Radware said the product includes a Risk Graph Map that continuously scores an organisation's security posture for agentic AI. The company said the map highlights multi-agent risk paths and potential data exposure in real time.

Standards alignment

Radware said the product aligns with the OWASP Top 10 for Agentic AI. It also uses the AI Vulnerability Scoring System, known as AIVSS, which it said it uses to assess and prioritise core security risks.

The company described the product as a security layer for autonomous workflows. It said it provides control over sensitive data and critical systems when AI agents operate across cloud services and internal applications.

ZombieAgent claim

The launch follows Radware's disclosure of a vulnerability it calls ZombieAgent. The company described it as a zero-click indirect prompt injection issue that affects agentic AI environments.

Radware said the flaw allows an attacker to implant persistent instructions into an AI agent's long-term memory or working context without user interaction. It said these instructions could trigger ongoing data exfiltration, including emails and corporate files, from cloud-based systems.

The company said such activity could bypass controls such as firewalls and endpoint detection tools. It framed the vulnerability as an example of a broader security visibility gap when agents interpret untrusted content.

"Vulnerabilities like ZombieAgent illustrate a critical structural weakness in today's agentic AI platforms," said Constance Stack, Chief Growth Officer, Radware.

"Enterprises rely on these agents to make decisions and access sensitive systems, but they lack visibility into how agents interpret untrusted content. Radware Agentic AI Protection was born to close this dangerous blind spot".

Wider portfolio

Agentic AI Protection sits alongside other Radware products aimed at AI-related risks. The company pointed to a recent launch of an LLM Firewall. It described that product as a way to address security concerns around large language model modules embedded in applications, including protection for prompts and responses against attacks and abuse.

Radware said Agentic AI Protection extends its broader agentic AI protection offering for enterprises, with further work planned around securing autonomous agents as adoption spreads across business functions.