From quarterly snapshots to continuous assurance: an agentic security advisor on Amazon Bedrock AgentCore
The customer
A Thailand-based cloud and data-privacy consultancy that advises its own clients on data-protection compliance, and holds internationally recognised certifications in information security management and privacy management.
AWS Partner: G-Able Public Company Limited
The challenge
The customer needed continuous, evidence-backed assurance over its AWS security posture, and could not get it economically by hand.
The requirement was sharper here than for a general enterprise: the company maintains certified management systems, so posture evidence is an obligation for its own surveillance audits — and it sells compliance services, so its credibility depends on its own posture being demonstrably sound.
Manual review of one AWS account against the CIS AWS Foundations Benchmark — roughly 62 controls — costs 4 to 5 senior man-days per round, and frameworks make that recurring: PCI DSS requires assessment at least quarterly, while CIS Controls v8 points toward continuous monitoring that periodic review cannot achieve.
Four constraints followed:
Cost of senior effort. For a consultancy, every man-day spent auditing its own estate is a man-day not delivered to a paying client.
Detection latency. With quarterly assessment, a misconfiguration introduced just after a round can sit undetected for up to 90 days. Most cloud data exposure originates from customer-side misconfiguration, and Thai PDPA penalties reach THB 5 million per case.
Evidence completeness. Manual review makes it hard to record consistently which controls were verified and which could not be reached — a distinction that matters when the output becomes audit evidence.
Turnaround. Answering one posture question meant waiting for a senior engineer to gather the evidence by hand.
Why this needed an agentic solution
“Is our data properly protected?” resolves into a chain — which buckets exist, how each is encrypted, which are public, what the key policies allow — and each answer changes the next question worth asking. A fixed rule engine cannot follow that chain, and a single model prompt has no access to live configuration.
The solution
G-Able delivered Splendormon, its AWS Cloud Security Posture Management (CSPM) and IAM security platform, as a dedicated single-tenant installation in an AWS account dedicated to this customer. The installation and all of the customer’s data at rest run in the Asia Pacific (Thailand) Region. The Amazon Bedrock AgentCore layer runs in Asia Pacific (Singapore), where the service is available, and the CloudFront web ACL is held in US East (N. Virginia), as AWS WAF requires for that scope.
At a high level
The platform continuously assesses the customer’s AWS estate against recognised benchmarks and exposes an AI security advisor that answers posture questions in Thai and English from live configuration evidence — not from generic knowledge. It reaches the customer’s accounts through a read-only role the customer installs themselves and can revoke at any time.
The technical detail
Genuinely agentic, not a prompt with a database. Every question goes to Claude on Amazon Bedrock with the schemas of all 39 tools. The model analyses first, then decides for itself whether to answer from knowledge or call tools, and which. Results feed back for a fresh round — up to 8 rounds within a 220-second per-turn budget — so the agent can retrieve evidence, reason about it, and decide to retrieve more before concluding. Tool choice is a model decision, not a routing rule.
All six AgentCore services in production. Runtime hosts the advisor agent; Gateway exposes the 39 tools over the Model Context Protocol; Identity provides workload identity and a token vault; Memory persists preferences and turn summaries for 90 days; Observability emits OpenTelemetry traces to AWS X-Ray; and Code Interpreter provides an isolated sandbox in which a proposed remediation script is validated before any human is asked to approve it — parsed for syntax, re-checked against the forbidden-action list, and screened for destructive commands and embedded credentials. The sandbox holds no AWS credentials by design and never executes the script against the estate; if it cannot be reached, the proposal fails rather than passing by default.
Agent authority is explicit, not implied. Permissions are enforced at three layers: user JWT authorisation; per-turn time and output-size budgets; and a tool allowlist enforced at both the application and the AgentCore Gateway, so a tool removed from the allowlist cannot be reached even if the model asks for it. Of the 39 tools, exactly one writes (report generation); the other 38 are read-only. A fourth gate sits on the remediation path specifically: no proposed script reaches a human approver until the Code Interpreter sandbox has validated it.
Least-privilege access. Splendormon audits target accounts through a cross-account read-only IAM role the customer installs from a G-Able CloudFormation template, protected by an ExternalId condition. Nothing is deployed into customer workloads, nothing is altered, and access is revoked by deleting the role.
Infrastructure. The platform runs on Amazon ECS on AWS Fargate with ARM64 tasks one per Availability Zone, behind an Application Load Balancer and Amazon CloudFront with an AWS WAF web ACL. State is held in Amazon RDS for MySQL Multi-AZ, with Amazon EFS for shared storage and interface VPC endpoints across both zones. The estate — 126 Terraform resources — is provisioned as code and deployed through GitHub Actions using OIDC federation with separated roles, no long-lived credentials.
Resilience proven by test. Multi-AZ is the production default, validated by fault injection against a live installation: 150 of 150 probes returned HTTP 200 while a running task was killed, and 238 of 240 during a forced database failover.
Results
| Measure | Before | After |
| Senior effort, the customer’s two-account estate assessed quarterly | ~40 man-days/year | ~4 man-days/year |
| Assurance model | Point-in-time, quarterly | Continuous |
| Maximum misconfiguration dwell time | Up to ~90 days | One scan cycle |
| Audit evidence | Assembled per round | Available on demand |
That is a saving of roughly 36 senior man-days per year, a reduction of approximately 90%. The account count and the quarterly cadence are the customer’s own; the per-account effort of 4 to 5 senior man-days is drawn from published market scoping for CIS AWS Foundations Benchmark assessment rather than measured on this customer.
The platform runs entirely on consumption-priced and On-Demand AWS services with no upfront cost or commitment, and the senior effort released substantially exceeds the cost of running it — on a model that deliberately values senior time at published government reference rates, well below commercial market pricing.
The shape of the work changed as much as the arithmetic. Evidence collection, control checking and report assembly leave the senior workload entirely, leaving risk prioritisation and remediation — the part that needs judgement.
Excluded from the return calculation: reduced breach exposure from shorter dwell time, and reduced regulatory exposure under the Thai PDPA.
About G-Able
G-Able Public Company Limited is a leading Thai technology enabler with over 37 years of experience, working with more than 100 global technology partners and serving over 1,000 active customers. G-Able’s AI practice spans cybersecurity, cloud and data centre modernization, data and analytics, digital business and application, and managed technology services, delivered by dedicated AWS-focused solution architecture, FinOps and delivery teams.
Want continuous assurance over your own AWS estate? [Contact G-Able].