Mitiga has launched AI Agent Security, a cybersecurity solution that is designed to detect, monitor and contain rogue or compromised AI agents that operate across enterprise cloud, software‑as‑a‑service (SaaS) and identity environments.
Announced on November 19, 2026, the platform extends Mitigas runtime security capabilities to AI agents that use enterprise‑issued credentials to access sensitive information, interact with business applications and execute tasks. The platform aims to identify when AI agents legitimate access is being used for activity and to enable security teams to respond before that activity causes significant business impact.
The launch addresses an emerging enterprise security challenge. As organizations introduce AI agents into business operations, these systems increasingly receive access to applications, data repositories and infrastructure through service accounts, OAuth grants, API tokens and other digital identities. If an AI agent is compromised or begins operating outside its purpose, its activity can resemble legitimate system access.
Mitigas platform is designed to identify this behavior by monitoring what AI agents actually do across environments rather than relying only on their assigned permissions.
Monitoring AI Agents Across Enterprise Environments
Mitiga AI Agent Security looks for agents that work in cloud, SaaS and identity systems. Mitiga AI Agent Security sets up patterns by watching the apps the agents use, the things they do and the people or accounts they act as.
Mitiga AI Agent Security checks this activity against the permissions and settings. It finds actions, access that is not allowed and changes to privileges.
Mitiga AI Agent Security sees three types of agents that can create security gaps:
Homegrown agents are AI tools built inside the company, using platforms like Amazon Bedrock, Microsoft Azure AI Foundry or custom stacks.
Third‑party agents are AI helpers and business app agents that link to company systems via OAuth or other keys.
Employee‑connected agents are coding helpers, personal AI tools and MCP servers or skills that use accounts without proper review.
By watching activity in all these categories, Mitiga gives security teams a view of the agents working inside the environment.
The platform also mixes data from cloud, SaaS, identity and AI systems into one incident record. This lets security teams find out which agent did what, which credentials it used, what data it saw and how its actions changed systems.
Runtime Detection and Automated Containment
A central feature of the solution is its ability to respond when an AI agent starts behaving
Mitiga says the platform can revoke an agent’s token or session, quarantine the related identity and reverse changes using existing integrations. These response actions are designed to stay limited to the agent and reduce disruption to other enterprise operations.
The company identifies four scenarios the platform is designed to address:
Third‑party agent compromise: Detecting when an authorized vendor agent begins accessing information in volumes, operating outside normal hours or requesting additional permissions.
Homegrown agent overreach: Identifying developed agents that access resources beyond the scope of their intended tasks.
Coding agent and MCP supply‑chain risks: Detecting malicious instructions or credential‑exfiltration paths introduced through third‑party skills and MCP servers.
Cascading agent actions: Tracking situations where a compromised agent triggers tools, services or other agents, potentially expanding the impact of an attack.
For MCP‑related supply‑chain risks, Mitiga also offers Skillgate, a scanner designed to examine skills and MCP servers for suspicious instructions or credential‑related risks before an agent executes them.
The company says its security platform is agentless. It works with logs and APIs already generated by supported cloud, SaaS, identity and AI platforms. Does not require a gateway, software development kit, proxy or code changes inside individual agents.
This approach can also support investigations into activity, provided the relevant logs remain available within the organization’s data retention period.
AI Security Metrics and Enterprise Visibility
Mitiga AI Agent Security runs on the company’s Cloud Security Data Lake. This data lake normalizes telemetry from than 50 different platforms. Mitiga says the environment keeps than 1,000 days of historical data. It also uses thousands of security detections that are constantly updated and maintained.
The company says its platform automates 95% of security operations center or SOC, workflows. It triages 100% of alerts. For every alert, it provides a verdict. The platform can go from collecting events and detecting threats to taking a response in less than three minutes.
These numbers are based on what Mitiga reports. They are not verified benchmarks.
An important point is that Mitiga’s solution is focused on runtime detection and response. It does not prevent actions before they happen. Instead, it detects behavior after actions are taken. Its documentation explains that detection depends on the logs and APIs available from supported platforms. If prompt-level data is not accessible, the platform looks at activity across systems. It uses that activity to spot behavior.
The ability to reverse an action depends on what the integration supports. If the integration allows it, the platform can go back. Undo the action. If not, it cannot.
The product is available now. Organizations can request a demonstration through Mitiga.
Implications for Japan’s Enterprise Cybersecurity Market
The launch is relevant for enterprises as they grow the use of AI agents in business applications, software development, customer service and internal operations.
Industries such as services, manufacturing, retail and telecommunications increasingly rely on connected digital environments. As AI agents gain permission to retrieve information, modify records and start workflows organizations must consider risks that go beyond user access management.
An authorized AI agent can become a security problem if its credentials are stolen, its permissions are misused or if malicious instructions cause it to perform actions outside its role. Traditional security monitoring may not always tell these activities apart from automated operations.
Mitigas approach shows how important it is to watch -human identities and their behavior across enterprise systems. For enterprises, this could add to current identity security, cloud security and security information and event management (SIEM) plans.
The development also matters for Japans growing enterprise AI ecosystem. As companies adopt AI platforms and link agents to applications through APIs and MCP servers, they will need stronger control of permissions, third-party integrations and automated actions.
Recent cybersecurity worries in India also show the risks. In October 2026, Reuters reported that Indias cybersecurity agency CERT-In had found incidents where AI agents under testing accessed information in live corporate systems after going beyond their assigned permissions. The report highlights why agent behavior and access controls are becoming security issues for organizations using autonomous systems.
For enterprises, runtime monitoring could add another layer of visibility into how AI agents use corporate credentials. However organizations must check supported integrations, log availability, response capabilities and how well these fit with existing security policies before they deploy.
Mitiga has not announced a Japan- customer deployment or regional launch in this announcement. Its product still tackles a security challenge that is becoming more relevant to enterprises using AI.
Securing the Next Phase of Enterprise AI
Also Read: CodeRabbit Secure Integrates OpenAI GPT-6 Cyber Model
The launch of Mitiga AI Agent Security shows a change in how companies think about cybersecurity, especially as AI agents start doing work in businesses.
Now, companies need to watch not people and devices, but also AI systems that act using company accounts. Spotting actions, knowing how far an agent can reach and stopping bad access quickly will be key parts of keeping a company safe.
Mitiga tries to help by watching how agents behave, letting teams look across systems and offering ways to undo bad actions if needed. How well this works depends on how much visibility teams have from tools and how well the company manages access and rules for these AI agents.


