Acalvio launches AI deception guardrails for agents

Acalvio has launched Deception Guardrails for AI agents, targeting a gap in AI security once an agent has been compromised.

The new offering embeds deceptive tripwires into agent workflows and surrounding AI infrastructure to detect and disrupt malicious activity before it reaches production systems. It is designed for organisations deploying autonomous AI agents that can reason, use tools and interact with application programming interfaces.

Most existing AI guardrail tools focus on filtering prompts and checking outputs. Acalvio argues those controls offer little help once an attacker has moved inside an agent's decision-making process or manipulated the infrastructure around it.

The launch comes as security vendors seek ways to address risks tied to agentic AI, a fast-growing category of software that can take actions with limited human oversight. Incidents involving AI agents have raised concerns that compromised systems could misuse credentials, call external tools or move across enterprise environments without triggering traditional controls.

Acalvio linked that risk to the recent Hugging Face incident, which drew wider attention to how quickly AI-related systems can be exploited. It also cited recommendations from cyber security groups including the Cloud Security Alliance, SANS and RSAC, which say deception can serve as an added defensive layer in environments where agents may struggle to distinguish real assets from decoys.

Ram Varadarajan outlined the company's position on the limits of existing controls.

"Reactive guardrails are designed to keep well-behaved AI systems on the road, but they do nothing to stop a hijacked agent driven by a malicious actor," said Ram Varadarajan, Chief Executive Officer, Acalvio.

He added: "With our patent-pending Deception Guardrails, we are moving the industry from reactive filtering to preemptive defense. If an AI agent goes rogue or its infrastructure is manipulated, our deceptive assets rapidly detect the misalignment, feed the attacker fabricated data, and alert the SOC before real enterprise assets are compromised."

How it works

The product extends Acalvio's ShadowPlex platform across on-premises and cloud environments. In practice, that means placing fake credentials, decoy services and other planted artefacts in areas where AI agents are likely to search for context or tools.

This includes what Acalvio calls agentic deception, in which honeytokens and decoy tools are inserted into files and configuration surfaces read by AI agents. If a compromised agent attempts to access or use those assets, the system generates alerts intended to give security teams early warning.

Acalvio also described a broader layer of decoy AI infrastructure. That includes decoy Model Context Protocol servers, decoy retrieval-augmented generation systems and decoy AI agents placed around the operational environment to draw out suspicious activity.

Another element is real-time monitoring for signs of manipulation, including jailbreak behaviour and prompt injection attempts. According to Acalvio, interactions with deceptive assets can expose those actions early enough to prevent movement into live enterprise systems.

Broader shift

The launch reflects a wider shift in cyber security as companies move from AI experimentation to larger-scale deployment. Security teams are increasingly being asked to protect not only data and networks, but also autonomous software agents that can read internal files, use credentials and connect to business systems.

That creates a different problem from earlier forms of AI governance, which focused largely on model behaviour, acceptable outputs and policy controls. In agentic settings, defenders must also consider what happens when a system gains access to tools, identities and infrastructure that can be abused if an attacker takes control.

Lawrence Pingree, Head of Research at Software Analyst Cyber Research, said the approach builds on existing deception methods already used in enterprise environments.

"Acalvio has spent years perfecting deception technologies to detect sophisticated attackers inside enterprise environments. Extending those same principles to AI agents is a natural and compelling evolution. By embedding deceptive assets directly into agent workflows and surrounding AI infrastructure with decoys, organizations gain a powerful new layer of detection that complements existing AI safety and governance controls," said Lawrence Pingree, Head of Research, Software Analyst Cyber Research.

The recommendations cited by Acalvio were blunt about why deception may work in this setting. They said agents are poorly placed to tell the difference between genuine credentials and honeypots, creating an opportunity for defenders to identify misuse through planted identities, package registries, datasets, application interfaces and other fake resources.

Acalvio said its Deception Guardrails are intended to provide full-spectrum coverage across enterprise environments, including internal systems and cloud infrastructure, so unauthorised lateral movement by AI agents can be spotted quickly.

The result, according to the company, is a higher-confidence signal for security operations teams than many conventional alerts, because interaction with a deceptive asset is itself a sign that something has gone wrong.