Safety & Security

Mitiga Launches Agentless Security Platform for Enterprise AI Agents

Mitiga launched AI Agent Security on November 19, 2026: an agentless platform that detects and contains suspicious AI agent activity from existing logs across cloud, SaaS and identity systems, triaging 100% of alerts in under three minutes.

Mitiga Launches Agentless Security for Enterprise AI Agents
Mitiga Launches Agentless Security for Enterprise AI AgentsLalmch / pixabay
By Marcus Bennett7 min read

Why it matters

  • Mitiga launched AI Agent Security on November 19, 2026, available immediately.
  • The platform triages 100% of alerts with an auditable verdict in under three minutes from event collection to response.
  • Its Cloud Security Data Lake normalizes telemetry from more than 50 platforms with over 1,000 days of retention.
  • CTO Ofer Maor said: "An agent doesn't need to break in when it already holds the credential. To the audit log, it looks like a valid login."
  • The free Skillgate scanner checks MCP servers and skills for hidden instructions or credential-exfiltration paths.

Mitiga launched AI Agent Security on November 19, 2026, an agentless platform designed to discover and monitor AI agents across cloud, software-as-a-service (SaaS) and identity systems without any code changes inside the agents themselves. The company made the platform immediately available, and it works from the logs and APIs that supported platforms already produce rather than from a gateway, SDK, proxy or any modification of the agent's execution path.

The stakes for enterprises are straightforward. AI agents can read data, call APIs, change records and trigger other agents using service accounts, OAuth grants (authorization delegations to third-party applications), API tokens, workload identities or delegated sessions. That combination of capability and credential access creates a visibility gap that traditional identity-based security tools do not address, because the agent never has to attack anything to act.

Why Existing Audit Tools Miss Agent Activity

Mitiga co-founder and chief technology officer Ofer Maor framed the core problem in a statement released with the launch: "An agent doesn't need to break in when it already holds the credential. To the audit log, it looks like a valid login."

That sentence describes the entire detection challenge the product targets. An agent operating with a legitimate OAuth grant or API token generates log entries indistinguishable from an approved user or service. Security teams reviewing audit logs see valid authentication events. The malicious or unauthorized behavior hides behind credentials that were, at issuance time, legitimate.

This is why Mitiga chose behavioral detection over inline blocking. The platform establishes behavioral baselines around three dimensions:

  • The systems agents access
  • The actions they perform
  • The identities they use

The platform then compares live activity against permissions and static configuration to surface privilege drift (gradual accumulation of access beyond an agent's intended scope) and unauthorized access. It correlates cloud, SaaS, identity and AI activity into a unified incident record, so investigators see one reconstructed event rather than fragments scattered across disconnected logs.

How Activity Flow Reconstructs Agent Behavior

A feature called Activity Flow maps users and non-human identities through AI platforms and connectors to the applications they ultimately reached. This gives security teams an end-to-end picture of what an agent touched, even when the agent traversed multiple intermediate systems.

Because Mitiga works from retained logs rather than from instrumentation inside agents, the platform can also support investigations into agent activity that occurred before deployment. If an enterprise suspects a compromised coding agent acted weeks ago, retained telemetry from cloud, SaaS and identity providers can be queried retrospectively. That historical capability distinguishes log-based detection from inline approaches, which only observe traffic that passes through them after installation.

What the Platform Covers

Mitiga defines coverage across three categories of agents, reflecting the messy reality of how AI tools enter enterprises:

  • Homegrown agents built on services such as Amazon Bedrock and Azure AI Foundry, developed by the enterprise's own teams
  • Sanctioned vendor agents connected through OAuth grants, approved but externally developed
  • Employee-connected tools, including coding assistants, personal AI tools and unreviewed MCP servers (Model Context Protocol servers, third-party extensions that give agents access to external tools and data) or skills

The third category deserves emphasis. Enterprises increasingly face agent sprawl that security teams never approved: a developer connects a coding assistant, an employee links a personal AI tool, someone installs an unreviewed MCP server. Mitiga's log-based approach can discover these agents precisely because it does not require the agent to register, install an SDK or route through a corporate gateway.

Mitiga names four threat scenarios the product addresses:

  1. Third-party agent compromise, where an externally developed agent is breached by an attacker
  2. Homegrown-agent overreach, where an internally built agent exceeds its intended permissions or scope
  3. Coding-agent and MCP supply-chain threats, where malicious instructions or data exfiltration paths are embedded in the tools and extensions agents consume
  4. Cascading actions, where one compromised agent triggers tools, services or other agents, propagating the compromise laterally

The cascading scenario reflects a structural property of agentic systems: agents that can trigger other agents multiply the blast radius of a single initial compromise.

Skillgate: A Free Scanner for the Supply Chain

For the coding-agent supply-chain risk specifically, Mitiga offers a free tool called Skillgate. The scanner checks skills and MCP servers for hidden instructions or credential-exfiltration paths before an agent acts on them. This is a pre-execution control, applied to the components an agent will consume rather than to the agent's runtime behavior.

Skillgate addresses a real gap. Skills and MCP servers are increasingly distributed through ecosystems with limited vetting, and a malicious skill can embed prompt-level instructions that direct an agent to send credentials to an external endpoint. Scanning those components before deployment catches the threat earlier in the chain than behavioral detection can.

If the agent has already run, Mitiga takes a different approach: the platform reconstructs what the developer's credentials did next, tracing the downstream actions that followed the malicious interaction.

Reported Operating Figures

Mitiga reports three operating figures for the platform's security operations:

  • 95% of security-operations-center workflows automated
  • 100% of alerts triaged with an auditable verdict
  • Under three minutes from event collection through triage to an actionable response

These numbers describe the detection and response pipeline, not the prevention of attacks. The "auditable verdict" framing matters for regulated enterprises, where security teams must document why an alert was closed or escalated.

The underlying architecture is Mitiga's Cloud Security Data Lake, which normalizes telemetry from more than 50 platforms and retains it for more than 1,000 days — nearly three years of searchable history. That retention window is what enables retrospective investigation of agent activity that predates the platform's deployment.

Where the Product's Limits Sit

Mitiga's own documentation is explicit about what the product is not: it is not an inline guardrail, and it does not block an action before that action occurs. Detection happens after the telemetry reaches the platform, and detection latency depends on the telemetry source.

Coverage is bounded by which logs exist. The product detects agents whose actions appear in supported cloud, SaaS, identity or AI sources. Agents operating entirely outside those sources fall outside the platform's view.

Prompt-level signals — the actual content of instructions sent to and from an agent — are available only where supported platform logs expose them. Mitiga names four such sources:

  • Bedrock logs
  • Azure OpenAI logs
  • Agentforce logs
  • Copilot logs

Where prompt-level visibility is absent, Mitiga falls back to detecting the resulting downstream behavior: the API calls, record changes and identity actions the agent performs, even if the platform cannot see what the agent was told.

What Containment Looks Like in Practice

When the platform identifies suspicious activity, responses run through customers' existing security integrations rather than through Mitiga itself. Supported actions include:

  • Revoking a token or session
  • Quarantining an identity
  • Reversing a change, where the specific integration supports reversal

Reversal capability depends on what each connected integration supports, which means containment options vary by customer environment. An integration that can revoke tokens cannot necessarily undo a record change an agent already committed.

Why This Launch Matters Now

For enterprises weighing agent security, the launch crystallizes a design trade-off the industry is still arguing about. Inline guardrails block actions before they occur but require agents to route through them, which misses unapproved and employee-connected tools. Agentless detection catches everything the logs show, including shadow agents, but reacts after the fact. Mitiga has chosen the second path, and its own documentation states the consequence plainly.

The practical implication for buyers is that log coverage and integration support form the platform's effective security boundary. Those two factors determine where Mitiga can investigate an agent's activity and which containment actions are available when it finds something. Enterprises evaluating the platform will need to audit their own telemetry sources first, because the answer to "are we covered" depends less on Mitiga than on what their cloud, SaaS and identity providers actually log — and for how long they keep it.

Source: unite.ai

Share this article:

More from Marcus Bennett

Marcus Bennett

Show full bio

Senior reporter covering companies, deals and strategy in artificial intelligence and machine learning at AI In Context.

791 articles

Related articles

  1. HYPR's AgentPass Ties Every AI Agent Action to a Named Human
  2. IBM Buys AuthMind to Watch What AI Agents Do After They Log In
  3. Tanium launches AgentSight to track and govern AI tools on endpoints
  4. Okta-Led Blueprint Alliance Wants a Kill Switch for Every AI Agent
  5. 83% of Firms Run Unapproved AI Coding Agents, BlueFlag Research Finds

« Previous articleNext article »