Tamnoon Builds an AI Cloud Security Remediation Agent on Amazon Bedrock

Cloud security teams have long faced an awkward reality: detection tools keep getting stronger, but remediation speed never keeps up. CNAPP (Cloud-Native Application Protection Platform) platforms spit out hundreds or thousands of security findings every day, yet only a fraction of them ever get fixed. Tamnoon is trying to break this bottleneck with AI agents—they built a cloud remediation orchestrator called Tami on Amazon Bedrock, letting AI autonomously investigate and safely remediate cloud security findings. According to official figures, this approach cut critical cloud exposure by over 90% within 90 days, with zero production incidents.

Why Cloud Security Remediation Became the Bottleneck

The traditional cloud security operations workflow goes roughly like this: CNAPP or cloud-native tools discover misconfigurations, excessive permissions, exposed ports, and other issues, generate alerts, and then security engineers investigate each one, confirm it, and fix it manually. The problem is the huge gap between automation on the detection side and manual work on the remediation side.

A typical mid-to-large cloud environment might run hundreds of AWS accounts, thousands of IaC templates, and microservices simultaneously. Every security finding requires understanding context—is this S3 bucket really exposed to the public? Is this IAM role depended on by other services? Manual remediation is not only slow but also risky: a single misconfigured security group rule can directly cause a service outage.

Tamnoon's entry point is exactly this "last mile." Rather than adding another layer of detection, they put AI agents into the remediation stage, letting the system complete investigation, decision-making, and execution on its own.

CNAPP is an integrated category proposed by Gartner, short for Cloud-Native Application Protection Platform. It consolidates several previously fragmented cloud security capabilities into a single platform: CWPP handles workload runtime protection, CSPM handles cloud misconfiguration and compliance baseline detection, and CIEM handles cloud identity and permission analysis, with some solutions also covering containers, Kubernetes, and IaC scanning. CNAPP's value lies in correlating attack surface, permissions, configuration, and runtime signals within the same context, enabling it to surface composite risks that single-point tools struggle to identify. But its output is still essentially a "findings list," and the volume of alerts scales linearly with cloud asset growth. When daily findings reach hundreds or thousands, human remediation capacity becomes the actual throughput ceiling of the entire security loop—precisely the bottleneck described in this article.

How Tami Works: Remediation Logic via Agent Orchestration

Tami is positioned as an "agentic cloud remediation orchestrator." Its core capability is not simply applying patches automatically based on rules, but simulating the reasoning process of a human security engineer.

When a security finding enters the system, Tami first conducts contextual investigation: which business unit does this resource belong to? What dependencies exist between it and other resources? How large is the blast radius of the remediation action? This information comes from CNAPP findings data, cloud-native APIs, and the environment graph Tamnoon has built up itself.

After completing the investigation, Tami generates a remediation plan and assesses the risk level. For low-risk, high-certainty issues, it can execute remediation autonomously; for changes involving critical production paths, it escalates to human review. The entire process runs on Amazon Bedrock, using foundation models for reasoning and decision-making.

The clever part of this design is the embedding of "safety guardrails." What Tamnoon emphasizes is not full automation, but "safe automation"—the agent must complete sufficient contextual verification before taking action, which explains how they achieved 90% remediation volume with zero production incidents.

IaC (Infrastructure as Code) refers to defining and managing cloud resources through declarative configuration files rather than manual console operations, with typical examples including Terraform, CloudFormation, and Pulumi. The key change it brings is that the true state of a cloud environment exists in two places simultaneously—the resources actually running in the cloud and the templates in the code repository. These two may diverge, which is known as "configuration drift." For security remediation, this means a single finding often requires fixing both the live resource and the IaC template, otherwise the next deployment will bring the problem back. This is also why "thousands of IaC templates" in mid-to-large environments significantly raises remediation complexity: the fix isn't just a cloud API call, but also tracing back to the corresponding code path while ensuring the pipeline isn't broken.

Why Choose Amazon Bedrock

Tamnoon chose Amazon Bedrock as its underlying platform for several practical reasons.

Bedrock provides unified access to multiple foundation models, which means Tami can select different models based on task type—the investigation phase may need stronger reasoning capabilities, while the remediation execution phase may prioritize speed and cost. Bedrock's managed nature also frees the Tamnoon team from building their own model inference infrastructure, allowing them to focus on logic orchestration in the cloud security domain.

From an architectural perspective, Bedrock's deep integration with AWS-native services is another advantage. Cloud security remediation itself requires extensive calls to AWS APIs, and building agents within the same ecosystem reduces integration friction.

Agentic orchestration is fundamentally different from traditional rule engines or workflow automation. Rule engines execute pre-written if-then logic, with coverage limited to what the author can foresee; agentic orchestration uses foundation models to reason at runtime, capable of calling tools, querying external data sources, and dynamically deciding the next action based on intermediate results. In cloud remediation scenarios, this capability manifests as: the agent can decide on its own which resources to query, how many layers of the dependency graph to expand, and whether a given change needs further verification. The trade-off is that behavior is not fully predictable, so constraint mechanisms must accompany it—tool whitelists, executable action boundaries, pre-change validation, risk tiering and human approval, and complete operation audit trails. The "safety guardrails" emphasized in the article refer precisely to these engineering measures that keep agent capabilities within acceptable bounds.

Lessons from Cutting 90% of Exposure in 90 Days

The data Tamnoon published deserves to be broken down: 90%+ reduction in critical exposure, a 90-day cycle, zero production issues. All three numbers are indispensable.

If there were only remediation speed without safety, the solution would be meaningless for production environments; if there were only safety without speed, its advantage over manual processes would not be obvious enough. Tamnoon's case shows that when AI agents complete sufficient contextual investigation before remediation and have clear risk tiering and human escalation mechanisms, automated remediation can remain controllable even under aggressive targets.

The lesson for cloud security practitioners is: the breakthrough point for the remediation bottleneck may not lie in better detection, but in letting AI act autonomously after understanding environmental context. The gap between detection and remediation is becoming one of the most valuable scenarios for AI agent deployment.

For security teams evaluating similar solutions, Tamnoon's practice offers a referenceable path—starting from the remediation stage, replacing item-by-item manual handling with agent orchestration, while treating safety guardrails as a first-class architectural element rather than an afterthought.

One-line Summary

Tamnoon uses its AI agent Tami on Bedrock to close the last mile of cloud security remediation, cutting exposure by 90% in 90 days.

Paragraph Summary

Cloud security has long suffered from a gap where detection is strong but remediation is weak: CNAPP generates large volumes of security findings daily, but limited manpower means only a small fraction are actually fixed. Tamnoon's entry point is the remediation stage, building an agent-driven cloud remediation orchestrator called Tami on Amazon Bedrock. Before acting, Tami conducts contextual investigation to determine resource ownership, dependencies, and blast radius, then performs risk tiering—autonomously executing low-risk, high-certainty issues while escalating those involving critical production paths to human review. Bedrock was chosen for its unified access to multiple models, managed inference infrastructure, and ease of integration with AWS-native services. Official data shows the solution cut critical cloud exposure by over 90% within 90 days with zero production incidents. The takeaway: the breakthrough for the remediation bottleneck may not be stronger detection, but enabling AI to act safely and autonomously after understanding environmental context.