Enkrypt AI, now part of Anaconda, has been named a Sample Vendor for Multimodal AI Protection in the Gartner® Emerging Tech Impact Radar: AI Cybersecurity Ecosystem.
As AI agents now take in images, texts, and voice, they can act on instructions hidden in any of them.
For example, a customer service agent receives a screenshot attached to a support ticket. Hidden in the image is an instruction to change the account’s payout details. The agent reads the image, follows the instruction, and calls the tool that updates the account. In a voice channel, a spoofed caller can do the same thing and get the agent to approve a wire transfer. Neither action leaves a log entry that would hold up in an audit.
What multimodal AI protection covers
In our red teaming work, we simulate adversarial scenarios across vision, text, and audio inputs. Gartner’s definition of the category covers those same input types:
“Multimodal AI protection secures AI systems across all data types and modalities that they process. Threats can come from many sources, including memory, context, external emails, documents used for grounding, and jailbreak instructions embedded in remote MCP configuration files. Attacks may be hidden in images, audio, video frames, pixel noise, metadata, or visual tricks. Security guardrails at test or runtime enable the uncovering of sophisticated threats that traditional security tools often miss.”
That definition covers both risks in this post: a poisoned screenshot, like the support ticket above, and the Model Context Protocol (MCP) servers agents connect to, covered below.
The radar places Multimodal AI Protection in its three-to-six-year ring, within the Securing AI (Runtime) segment. The report explains that “‘Range’ represents Gartner’s estimate of the time it will take each supporting technology to reach early majority adoption (more than 15% target market adoption), not when product leaders should act on an investment.”
What multimodal attacks look like in practice
Our red teaming also simulates cross-modal injection attacks, where visual content is used to covertly encode malicious instructions.
In our May 2025 multimodal red teaming of two vision-language models, one technique placed a harmful request inside an image, as the title of an empty numbered list, then sent a text prompt asking the model to fill in each item. The text prompt looked harmless on its own. Across all the harmful prompts we tested on both models, 68% produced unsafe content.
A text-only filter misses that attack because it never inspects the image. The same pattern shows up in other forms:
- Image-and-text prompt smuggling, where an image carries the instruction
- Audio injection and transcription manipulation
- Cross-modal chains, such as image to text to tool call, or audio to text to tool call
- Retrieval and memory poisoning, where a document or stored context carries the instruction
In each case, the input carries an instruction the model may follow, and a tool call can turn that instruction into an action.
Where Enkrypt AI fits
Our customers tell us securing AI agents is uniquely challenging, because every interaction has to stay safe, secure, and compliant. Enkrypt AI tests multimodal systems before they ship, then applies the same policy to text, image, and audio inputs in production.
| What the job requires | How Enkrypt AI handles it |
|---|---|
| Test before release | Agent Red Teaming tests models and agents across 300+ attack categories, using adversarial image-text pairs, visual perturbations, and audio distortions for multimodal systems. |
| Inspect inputs at runtime | Agent Guardrails apply policies to text, vision, and audio, with defenses against image-and-text prompt smuggling, audio injection, and cross-modal chains such as image to text to tool call. MCP Gateway enforces those checks inline on MCP tool calls. |
| Enforce one policy across modalities | Agent Policy Engine turns internal policy and regulation into one set of controls and generates audit-ready evidence. Every guardrail decision is logged with a policy ID, version, and reason code. |
For the support ticket above, Enkrypt AI Agent Guardrails approve or deny the account update before it runs. They deny the action even if the model attempts it, so a hidden instruction does not become a wire transfer or an account change.
Together, these capabilities help security teams find multimodal risks before an AI agent ships, stop hidden instructions while it runs, and show compliance reviewers which policies were enforced in production.
See how Enkrypt AI secures multimodal AI agents.
One policy across testing, runtime, and MCP
Multimodal protection runs inside AI security & guardrails in the Anaconda Platform. It helps teams test AI before release, protect it at runtime, govern tool and data access, and prove compliance, all governed by one policy engine.
We consider that MCP is one of the threat sources, and in our experience the exposure is already widespread. In two months, Enkrypt AI scanned 268,000+ tools across 25,000 MCP servers with its MCP Scanner, uncovering 143,000+ vulnerabilities affecting 73% of those servers.
Enkrypt AI’s recent Gartner recognitions:
- September 2025: Named a Gartner Cool Vendor in AI Security.
- September 2026: Named a Specialist in the first Gartner Emerging Market Quadrant for AI Application Security — Startup Vendors.
- October 2026: Named a Sample Vendor for Multimodal AI Protection in the Gartner Emerging Tech Impact Radar: AI Cybersecurity Ecosystem.
Where to start
List every place your AI systems accept something other than typed text: uploaded files, images, voice, retrieved documents, and MCP tool responses. Then red team one of those paths with multimodal attacks before it reaches production. The results show which guardrails you need and give your security and compliance reviewers concrete evidence to sign off on.
See Enkrypt AI in action. Request a demo with one of your own multimodal agents and see what red teaming finds and what guardrails stop.
Gartner clients can read the full Emerging Tech Impact Radar: AI Cybersecurity Ecosystem on gartner.com.
FAQs
What is the Gartner Emerging Tech Impact Radar?
In Gartner’s words, “This Emerging Tech Impact Radar analyzes and illustrates two significant aspects of impact: when we expect a technology to have a significant impact on the market (namely, the range); and how big of an impact it has on relevant markets (specifically, mass).” This edition covers 22 emerging technologies and trends in the AI cybersecurity ecosystem.
What did Gartner name Enkrypt AI by Anaconda?
In its Emerging Tech Impact Radar: AI Cybersecurity Ecosystem, published October 5, 2026, Gartner named Enkrypt AI (now part of Anaconda) a Sample Vendor for Multimodal AI Protection.
Has Gartner recognized Enkrypt AI before?
Yes. Gartner named Enkrypt AI a Cool Vendor in AI Security in September 2025, and a Specialist in the Emerging Market Quadrant for AI Application Security — Startup Vendors in September 2026.
Why do multimodal AI systems need their own protection?
A multimodal model treats an image, a document, or an audio clip as input it can act on. An instruction hidden in any of them can steer the model the same way a typed prompt can, and text-only filters do not inspect those inputs.
How does Enkrypt AI relate to Anaconda?
Enkrypt AI is part of Anaconda. Enkrypt AI by Anaconda delivers AI security & guardrails in the Anaconda Platform, one of four platform pillars alongside AI workspaces, AI orchestration, and AI artifacts.
Gartner, Emerging Tech Impact Radar: AI Cybersecurity Ecosystem, David Senf, Mark Wah, Tarun Rohilla, Marissa Schmidt, Evan Zeng, Alfredo Ramirez IV, Tuong Nguyen, Walker Black, 5 October 2026, ID G00846953.
Gartner, Cool Vendors in AI Security, Jeremy D’Hoinne, Bart Willemsen, Dennis Xu, Avivah Litan, 24 September 2025.
Gartner, Emerging Market Quadrant for AI Application Security — Startup Vendors, Meghan Hollis, Dionisio Zumerle, Dennis Xu, Marissa Schmidt, 16 September 2026, ID G00856640.
GARTNER is a trademark of Gartner, Inc. and/or its affiliates.Gartner does not endorse any company, vendor, product or service depicted in its publications, and does not advise technology users to select only those vendors with the highest ratings or other designation. Gartner publications consist of the opinions of Gartner’s business and technology insights organization and should not be construed as statements of fact. Gartner disclaims all warranties, expressed or implied, with respect to this publication, including any warranties of merchantability or fitness for a particular purpose.