Inside the AI agent that fixes cloud findings without touching production blind

The gap between cloud threat detection and resolution is widening. According to the 2026 State of Cloud Remediation report, 53% of detections remain open, up from 41% the previous year. While Cloud Native Application Protection Platforms (CNAPPs) provide visibility, operational capacity has not kept pace. Critical alerts now stay open for an average of 150 days, a steep climb from under 40 days in 2024.

Tamnoon addresses this backlog through Tami, an AI cloud security agent. Rather than a single chatbot, Tami is a multi agent system governed by specialized skills and a human supervised safety framework.

The architecture of specialized skills

AI in cloud operations often suffers from a hallucination of authority, where an agent suggests disruptive changes like a terraform destroy because it lacks environmental context. Tami mitigates this by decomposing remediation into discrete skills.

Instead of one model solving a misconfiguration, Tamnoon uses multiple agents to answer specific operational questions:

• Asset identification: Defining the resource and its business function.
• Dependency mapping: Identifying services that will break if a resource is modified.
• Ownership discovery: Finding the responsible engineer based on activity signals rather than stale tags.
• Historical context: Reviewing successful fixes in similar environments.
• Risk evaluation: Calculating the likelihood of a production outage.

This modularity allows for higher precision. By isolating logic, the system applies specific guardrails to each step. The ownership skill, for instance, uses layered signals to find the right contact, bypassing the inaccurate tagging that often stalls remediation.

Quantifying risk: the remediation confidence score

The fear of downtime is the primary barrier to automated remediation. Tamnoon uses a Remediation Confidence Score to categorize fixes into three states:

1. SAFE: The fix has a negligible blast radius. Dependencies are understood, and actions like enabling encryption on an unused bucket are unlikely to cause disruption.
2. RISKY: The fix involves critical path infrastructure or complex dependencies. These require manual execution during maintenance windows.
3. AWAITING DATA: The system lacks enough telemetry to judge the impact.

This scoring system supports the claim of zero production incidents across every remediation Tami has executed. By flagging what it does not know, the system prevents the automated errors that engineers fear.

Human in the loop: CloudPros

Complex enterprise environments present edge cases that defy pure automation. Tamnoon incorporates CloudPros, human cloud security experts who act as a final verification layer.

When Tami generates a remediation plan (including CLI commands, IaC templates, and validation scripts), a CloudPro reviews the logic and safety parameters. This hybrid approach ensures the output is a deployable piece of code that fits within existing change management processes like Jira or ServiceNow. The system handles 90% of the manual effort, leaving the final decision to a human.

Ecosystem integration

Tamnoon functions as a tool agnostic copilot, integrating with Wiz, Orca Security, Prisma Cloud, and Microsoft Defender for Cloud.

A recent technical development is the introduction of a Model Context Protocol (MCP) server for Cortex Cloud. This allows AI agents to consume context (such as ownership data and fix history) directly from the Palo Alto Networks ecosystem. Tamnoon has also joined Anthropic's Cyber Verification Program (CVP) to test these workflows against safety standards.

The economic case for remediation

The scale of the problem is visible in alert volumes. In 2025, the High severity queue was 17 times the size of the Critical queue. While teams focus on criticals, the mass of high severity debt creates a long tail of risk.

For organizations struggling with this backlog, hiring more headcount is often unsustainable. One Fortune 1000 healthcare company achieved an 87% reduction in cost per remediation by shifting from manual cycles to this managed model.

As vulnerability closure times slow (rising to 282 days in 2026), the transition from detection to supervised remediation is a requirement for cloud scale operations.