Traditional security guardrails currently focus on filtering inputs and outputs, leaving security operations centers blind once an attacker gains control of an agent. Acalvio’s new capability, which extends the company's ShadowPlex platform, shifts this dynamic by surrounding AI infrastructure with a deceptive reality. By deploying honeytokens, decoy RAG systems, and fake Model Context Protocol servers, the technology forces rogue agents to reveal their malicious intent the moment they interact with non-production assets.

Ram Varadarajan, CEO at Acalvio, noted that reactive filtering fails to stop a hijacked agent driven by a malicious actor. This preemptive approach aligns with industry recommendations, such as those from the Cloud Security Alliance, which advocate for deception as a primary defense because agents cannot inherently distinguish between valid production credentials and honeypots. By embedding these traps directly into agent workflows, organizations can generate high-confidence alerts and neutralize threats—including prompt injections and unauthorized lateral movement—before business-critical data is compromised.