In an AWS blog article, Tamnoon describes Tami, an agentic cloud remediation orchestrator built on Amazon Bedrock. It handles security findings across AWS, Azure, Google Cloud and other providers. The described workflow separates investigation from changes and requires an independent safety assessment before remediation is drafted.

From findings to decisions

The platform ingests findings through API integrations with CNAPPs, CSPM providers and cloud-native tools. It normalizes them into a unified schema, then adds asset and exposure context before prioritization. Named sources include Wiz, CrowdStrike, Palo Alto Cortex, Upwind, SentinelOne, Orca, Cyera, AWS Security Hub and GuardDuty.

Tami delegates work to four specialized agents. The investigation agent produces read-only steps and a complete plan, supported by an LLM-as-a-judge validation layer. The safety-scoring agent combines automated verification with analyst findings into an evidence chain for each asset. With a conservative bias, it assigns SAFE, RISKY or AWAITING DATA. This Remediation Confidence Indicator distinguishes work eligible for automation from work requiring human review.

Preparing and reviewing changes

The remediation agent activates only for findings deemed safe. It plans ordering, dependencies and conditions, with each step structured around backup, remediation and verification, plus rollback guidance. The code-generation agent converts the plan into executable remediation after an input guardrail screens relevance and prompt injection. Every artifact passes through a generation and review loop requiring approval from a separate reviewer agent.

According to the article, outputs are checked against automation results, user feedback and training data from actual resolved alerts. The process includes reflection, regeneration and human involvement when needed. AI capabilities undergo internal testing before customer use.

Models and infrastructure

Tamnoon selected Bedrock for access to models through one API, allowing model choices to reflect the quality and cost needs of reasoning, validation and review. The described Tami workloads invoke Anthropic Claude through ChatBedrockConverse. Inference runs in Tamnoon's AWS account in US commercial Regions, with Cross-Region inference routing within the US. The article states that Bedrock does not retain prompts or completions or use them for training.

The platform operates in Tamnoon's own AWS accounts, isolated by environment. API Gateway provides entry, Certificate Manager manages encryption and Cognito handles sign-in. The core runs in private networks across multiple Availability Zones. EKS hosts dashboard services, while Lambda collects and analyzes findings and executes remediations. EventBridge, SQS and Step Functions coordinate parallel processing so a bad record does not stall the rest.

Permissions, data and oversight

Bedrock AgentCore governs agent actions, and Guardrails review requests and generated code. The described controls cover every request and response, mask PII and secret-like content, filter prompt injection and log interventions. Lambda services run in private VPC subnets without public ingress and with controlled egress.

AI agents never hold cloud credentials. Isolated execution uses short-lived, tenant-scoped credentials under customer-controlled trust. Each service has a dedicated IAM role and restricted permissions; Bedrock access allows specified model invocations without model management, training or export permissions. Each invocation processes one SQS message and one tenant's data. Service authentication uses application-scoped OAuth2 client credentials through Cognito.

Managed databases and caching hold state, S3 stores playbooks and artifacts, and Secrets Manager holds credentials. Outcomes can flow to Slack, Jira, ServiceNow and GitHub. Remediations pass through approval workflows with audit trails. Configuration is versioned in CloudFormation and SAM, and models are pinned to dated identifiers. The article reports CloudTrail records for every inference call and a decoding temperature of zero. Queues, error handling and CloudWatch alarms connected to on-call support operational oversight.

Reported results and their limits

The article says Tami has analyzed more than 14 million cloud alerts. Tamnoon reports a reduction of over 90% in critical cloud exposure within 90 days, with zero production issues. These outcomes are attributed to its own research; the supplied evidence does not establish independent validation, sample size or measurement methodology. Claims of less manual work and more exposures closed per analyst are not quantified. SAFE actions can execute automatically, while RISKY and AWAITING DATA actions involve Tamnoon's embedded cloud engineers. The reported outcomes are not guarantees for every customer.