Agentic Defense
Runtime control for the prompts, tool calls, and identities that AI agents use — on infrastructure you control.
The last month made a shift plain. Attackers are no longer only using chatbots to draft better phishing mail. They are fielding autonomous agents, abusing the interfaces that connect assistants to tools, and stealing the credentials and transcripts that AI systems leave behind.
On 16 November 2026, Google Threat Intelligence Group and Mandiant reported that agentic threat activity had crossed a significant threshold. Open-source multi-agent frameworks are now appearing in live intrusions, and GTIG said it had observed, for the first time, a pipeline that could discover and exploit vulnerabilities without a human in the loop. Two weeks earlier, Microsoft described how China-aligned Storm-0940 used the ARTEX agentic framework to attempt sign-ins to more than 2,250 Microsoft 365 tenants after unauthorized access associated with ConnectWise Cloud Backup, with some environments reaching privileged access in under three hours. In Germany, hosting provider Mittwald disclosed that a vulnerability in its Model Context Protocol (MCP) server — the interface customers use to connect AI assistants to hosting accounts — was exploited on 27 October and affected 56 customer projects. In the same window, breaches at AI infrastructure provider Fireworks AI exposed customer billing data and, in downstream cases, Notion AI chat transcripts and Cursor coding-agent traces. Some of those traces held unredacted credentials that were later used to plant a malicious coding-assistant extension in Cursor’s marketplace. On 21 October, a compromised LiteLLM release shipped credential-stealing malware to PyPI and container registries. And on 17 November, Bitdefender Labs reported a GitHub campaign that plants malicious instructions, session hooks, and MCP configurations to steer AI coding assistants into running hidden payloads.
The methods differ. The pattern does not: once an agent can call tools, read mail or code, or hold credentials, a poisoned instruction, a compromised connector, or a leaked trace can move faster than a human analyst.
ZeroTrusted.ai was built for that pattern. The AI Firewall sits between users, models, and tools and inspects every prompt, response, and tool call. Policies detect prompt injection and jailbreaks, redact or block sensitive data before it leaves your perimeter, allow-list the tools an agent may invoke, inspect arguments, and revoke tool access when a policy is violated. Runtime enforcement and identity policy give each agent an owner, permissions, and an autonomy limit, so a confused or compromised agent cannot silently inherit standing access to everything a user can touch.
When something still needs a response, AI SOAR puts 300+ agents, 200+ connectors, and 300+ built-in tools to work correlating signals, investigating, containing, and retesting — with a person required to approve high-impact actions. Every prompt, response, tool call, and approval can be traced. Gartner’s “Operationalize AI Governance With Ongoing Postdeployment Control” (G00837065, 21 October 2025) names ZeroTrusted.ai, among others, as offering runtime enforcement. ZeroTrusted.ai was also a Representative Vendor in Gartner’s “Innovation Guide for Generative AI in Trust, Risk and Security Management” (G00799579, September 2024).
The platform is on-premises first. It supports private-cloud and air-gapped installs on STIGed Kubernetes, and is currently set to FedRAMP High for Microsoft Azure and AWS. ZeroTrusted.ai does not claim a FedRAMP authorization; authorization stays with the customer’s system and assessor. The platform supports SOC 2 and other packaged control mappings to help you produce evidence.
Bring the models, agents, and MCP servers you already use. We add security, privacy, and transparency — from a forward-deployed node to the enterprise data center.
