Discovering AI agents in AWS accounts with Amazon Bedrock for a cybersecurity startup
IsaMar Technology Inc. helped a cybersecurity software startup find autonomous, tool-using AI agents in AWS accounts from cloud telemetry and configuration, separating them from plain model usage. A serverless service filters AWS CloudTrail activity in code and passes the candidate activity to a bou
AWS Partner: IsaMar Technology Inc. Customer: An early-stage cybersecurity software company in the SMB segment Status: In production on AWS as of November 2026 AWS services: Amazon Bedrock (Anthropic Claude), Amazon Bedrock Guardrails, Amazon Bedrock AgentCore, AWS Lambda, AWS CloudTrail, Amazon S3, AWS KMS, AWS IAM, AWS IAM Identity Center, Amazon EventBridge, Amazon CloudWatch, AWS Security Hub
About the customer
The customer is an early-stage independent software vendor (ISV) in cybersecurity, in the small and medium business segment. It builds security software for organizations that run AI agents, sells to software companies and enterprises, and runs its own engineering on AWS. The customer asked not to be named.
The challenge
AI agents can be built and run in an AWS account without the security team knowing. The customer wanted to find them from the account itself: autonomous, tool-using, multi-step agents, as distinct from applications that simply call a model. It had no account-level view of agent activity, no pipeline for cloud audit logs, and no measure of what an agent's IAM role could actually reach. An agent that had been configured but had not yet run left no activity to see at all.
Calling an agent "unapproved" needed ground truth, and no formal list of approved agents existed. The monitored account's telemetry, prompts and verdicts had to stay in that account, and detection had to be measured against known answers and people's reviews, not assumed.
The solution
IsaMar Technology Inc. delivered the solution in production in the customer's AWS account: a serverless discovery service, a bounded investigation agent on Amazon Bedrock, and scheduled discovery that runs them on the account's own activity.
In the account, with no standing compute. A dedicated AWS CloudTrail trail delivers management, global service and Amazon Bedrock data events, with log file validation, to its own Amazon S3 bucket. One AWS Lambda function reads a bounded time window, groups events into sessions and applies deterministic checks in code, so only activity that shows signs of agency reaches a model. Through a read-only role it also reads Amazon Bedrock Agents and Amazon Bedrock AgentCore configuration, which finds agents that are configured but have not yet run. Lambda fits a bounded batch run better than a streaming tier on Amazon Kinesis and AWS Fargate, and keeps telemetry inside the account.
An investigation agent on Amazon Bedrock, in production. For each candidate, a bounded investigation agent built with the Strands Agents SDK, running Anthropic Claude Sonnet 4.5 on Amazon Bedrock through a tagged application inference profile over a US cross-Region inference profile, plans its investigation first, then calls read-only tools for the cited events, the agent's configuration, what its IAM role can reach and the approved list, and records a structured verdict through a forced tool call. Code checks every cited event, and an invented citation marks the evidence unsound. SDK limits and hooks outside the model's reach cap turns, tool calls, tokens and time, and any stop falls back to a single classification call at temperature 0, flagged for review. The agent can read and reason, and it cannot change anything in the account.
Responsible AI controls. Prompts carry event metadata, never request or response bodies. Whether an agent is approved is a lookup in code against the customer's list, and the model is never asked. Amazon Bedrock Guardrails screens every model call, with a prompt-attack filter on input, content filters on output, and AWS keys and passwords blocked both ways, because names a workload gives itself are a prompt-injection path. Doubtful verdicts and unapproved agents go to a person, whose recorded review is the ground truth for the measures below and feeds an append-only refinement log. A known-answer suite of planted test agents, including false-positive traps, runs in a separate test account and measures detection, and an untested trap counts against the result.
Scheduled discovery. An Amazon EventBridge rule runs the same function on a schedule, in production. Each scan reads the activity delivered since the last one, records one finding per IAM role, and holds its place on any failure so the next scan picks it up; a daily token budget caps what the scans may spend, and an alarm treats a silent schedule as a failure, not a quiet account. Findings sit in their own encrypted Amazon S3 bucket that refuses deletes, apart from the planted test agents.
Security and operations. The pipeline role can read trail content and call Amazon Bedrock only from inside the function, and the agent gains no permission of its own. Every role sits under a permissions boundary, with no IAM users or long-lived keys: engineers sign in with MFA through an AWS IAM Identity Center permission set, then assume a scoped role with an external ID. Results stay in Amazon S3 stores encrypted with AWS KMS, and only aggregate counts and rates leave the account. Amazon CloudWatch alarms watch each scan, and a scan whose reads or model calls failed holds rather than reporting clean. Unapproved agents can become AWS Security Hub findings. Infrastructure is HashiCorp Terraform with automated plan tests.
Results
The solution is in production on AWS as of November 2026.
- Account-level discovery: from no account-level view of agent activity to scheduled discovery in production in one monitored AWS account, from the account's own telemetry.
- Agents that have not run: Amazon Bedrock Agents and Amazon Bedrock AgentCore configuration is read on every scan, so a configured agent is found before it ever runs.
- Evidence a reviewer can check: code checks every event a verdict cites, and every investigated verdict carries the agent's plan and the tools it relied on.
- Detection on known answers: 95% precision and 95% recall (target at least 90%; 19 of 20 known-answer cases on the first pass; the one miss was caught on re-run), run in the separate test account.
- Agents found outside the existing inventory: from 0 to 7 (target at least 5).
- Time from agent activity to finding: 12 minutes (target at most 15 minutes), where none could be found before.
About IsaMar Technology Inc.
IsaMar Technology Inc. is an AWS Advanced Tier Services Partner, founded in 2025, delivering AWS cloud and DevOps, custom software, and CRM and ERP work for enterprises, government agencies and prime contractors nationwide. It is a WBENC-certified Women's Business Enterprise and an NMSDC-certified Minority Business Enterprise.
Customer name withheld at the customer's request. IsaMar Technology Inc. is an AWS Advanced Tier Services Partner.
Need a certified diverse partner on your next technology program?
Tell us about the contract, the scope, and your supplier diversity goals. We’ll respond with how we can deliver and how the spend can be counted.
Start a conversation Supplier diversity details