The dialog round AI and its enterprise purposes has quickly shifted focus to AI brokers—autonomous AI techniques that aren’t solely able to conversing, but additionally reasoning, planning, and executing autonomous actions.
Our Cisco AI Readiness Index 2025 underscores this pleasure, as 83% of corporations surveyed already intend to develop or deploy AI brokers throughout a wide range of use instances. On the identical time, these companies are clear about their sensible challenges: infrastructure limitations, workforce planning gaps, and naturally, safety.
At a time limit the place many safety groups are nonetheless contending with AI safety at a excessive stage, brokers broaden the AI threat floor even additional. In any case, a chatbot can say one thing dangerous, however an AI agent can do one thing dangerous.
We launched Cisco AI Protection in the beginning of this yr as our reply to AI threat—a very complete safety resolution for the event and deployment of enterprise AI purposes. As this threat floor grows, we wish to spotlight how AI Protection has advanced to satisfy these challenges head-on with AI provide chain scanning and purpose-built runtime protections for AI brokers.
Under, we’ll share actual examples of AI provide chain and agent vulnerabilities, unpack their potential implications for enterprise purposes, and share how AI Protection permits companies to immediately mitigate these dangers.
Figuring out vulnerabilities in your AI provide chain
Fashionable AI growth depends on a myriad of third-party and open-source elements reminiscent of fashions and datasets. With the appearance of AI brokers, that record has grown to incorporate property like MCP servers, instruments, and extra.
Whereas they make AI growth extra accessible and environment friendly than ever, third-party AI property introduce threat. A compromised part within the provide chain successfully undermines the whole system, creating alternatives for code execution, delicate information exfiltration, and different insecure outcomes.
This isn’t simply theoretical, both. A number of months in the past, researchers at Koi Safety recognized the primary identified malicious MCP server within the wild. This package deal, which had already garnered 1000’s of downloads, included malicious code to discreetly BCC an unsanctioned third-party on each single e mail. Related malicious inclusions have been present in open-source fashions, device recordsdata, and numerous different AI property.
Cisco AI Protection will immediately deal with AI provide chain threat by scanning mannequin recordsdata and MCP servers in enterprise repositories to determine and flag potential vulnerabilities.
By surfacing potential points like mannequin manipulation, arbitrary code execution, information exfiltration, and gear compromise, our resolution helps stop AI builders from constructing with insecure elements. By integrating provide chain scanning tightly throughout the growth lifecycle, companies can construct and deploy AI purposes on a dependable and safe basis.
Safeguarding AI brokers with purpose-built protections
A manufacturing AI software is inclined to any variety of explicitly malicious assaults or unintentionally dangerous outcomes—immediate injections, information leakage, toxicity, denial of service, and extra.
Once we launched Cisco AI Protection, our runtime safety guardrails have been particularly designed to guard towards these situations. Bi-directional inspection and filtering prevented dangerous content material from each person prompts and mannequin responses, protecting interactions with enterprise AI purposes secure and safe.
With agentic AI and the introduction of multi-agent techniques, there are new vectors to contemplate: higher entry to delicate information, autonomous decision-making, and sophisticated interactions between human customers, brokers, and instruments.
To satisfy this rising threat, Cisco AI Protection has advanced with purpose-built runtime safety for brokers. AI Protection will operate as a form of MCP gateway, intercepting calls between an agent and MCP server to fight new threats like device compromise.
Let’s drill into an instance to higher perceive it. Think about a device which brokers leverage to look and summarize content material on the internet. One of many web sites searched comprises discreet directions to hijack the AI, a well-known state of affairs often known as an “oblique immediate injection.”
With easy AI chatbots, oblique immediate injections would possibly unfold misinformation, elicit a dangerous response, or distribute a phishing hyperlink. With brokers, the potential grows—the immediate would possibly instruct the AI to steal delicate information, distribute malicious emails, or hijack a related device.
Cisco AI Protection will shield these agentic interactions on two fronts. Our beforehand present AI guardrails will monitor interactions between the appliance and mannequin, simply as they’ve since day one. Our new, purpose-built agentic guardrails will study interactions between the mannequin and MCP server to make sure that these too are secure and safe.
Our objective with these new capabilities is unchanged—we wish to allow companies to deploy and innovate with AI confidently and with out worry. Cisco stays on the forefront of AI safety analysis, collaborating with AI requirements our bodies, main enterprises, and even partnering with Hugging Face to scan each public file uploaded to the world’s largest AI repository. Combining this experience with a long time of Cisco’s networking management, AI Protection delivers an AI safety resolution that’s complete and completed at a community stage.
For these concerned with MCP safety, take a look at an open-source model of our MCP Scanner you can get began with at present. Enterprises on the lookout for a extra complete resolution to deal with their AI and agentic safety issues ought to schedule time with an knowledgeable from our crew.
Lots of the merchandise and options described herein stay in various phases of growth and shall be provided on a when-and-if-available foundation.
