A frontier AI agent is now doing design review, threat modeling, code review, and penetration testing — as one unified service. Here's how it works and how to roll it out.
Most security teams don't get to choose whether they have enough time to test everything — they choose which applications get tested this quarter, and live with the risk on everything else. AWS Security Agent, now folded into the broader AWS Continuum suite, is AWS's attempt to remove that trade-off by turning security work into an on-demand, agentic capability rather than a scheduled, expert-gated one.
This is a full walkthrough: what the service actually is, the specific problems it targets, how its four capabilities fit together, and a concrete path to implementing it.
Previewed at re:Invent 2025 and expanded steadily through 2026, AWS classifies Security Agent as a frontier agent: an autonomous system that works independently toward a goal, scales to handle many tasks concurrently, and can run for hours or days without constant supervision. That's a meaningfully different operating model from a conventional AI assistant that just responds to prompts one at a time.
In practice, the agent ingests an application's design documents, source code, and architecture context, builds an internal model of how the system fits together — its components, data flows, and trust boundaries — and uses that understanding to perform four connected functions:
Because all four share the same underlying reasoning about the application, findings from one stage inform the others — a threat model can flag the trust boundaries a subsequent penetration test should target first.
Manual pentesting is expensive and slow, so most orgs only test their most critical apps, and only periodically — leaving most of the portfolio unassessed between cycles. Security Agent turns this into an on-demand, 24/7 capability.
Rigorous threat modeling traditionally needs a security architect manually mapping data flows and trust boundaries — often skipped under deadline pressure. The agent automates this directly from design documents or code.
Static analysis tools match known vulnerability signatures — effective against known patterns, blind to issues that emerge from how components interact. Security Agent reasons about architecture and data flow instead.
Security review is often a late, manual gate before release. Embedding review, threat modeling, and validated remediation into pull requests and IDEs lets teams fix issues inline, without a context switch.
Teams are often flooded with theoretical findings. Security Agent validates findings in simulated environments to demonstrate real exploitability, with CVSS and app-specific severity ratings to prioritize.
Design review, threat modeling, code review, and pentesting are traditionally separate tools and vendors. Security Agent unifies design-time, development-time, and deployment-time security into one offering.
Continuously validates architecture and design docs against managed compliance packs — AWS WAF, NIST CSF, and PCI DSS — or your own requirements imported from internal docs or Confluence. Every finding maps back to compliance posture.
From design documents or a connected repo, the agent builds a contextual model — components, data flows, architecture, trust boundaries — then identifies threat actors and attack vectors and prioritizes across all six STRIDE categories (Spoofing, Tampering, Repudiation, Information disclosure, Denial of service, Elevation of privilege).
Supports per-PR and full-repository scanning across GitHub, GitLab, and Bitbucket (SaaS and self-hosted), with Confluence integration for context. It reasons about trust boundaries and data flow rather than matching signatures, then generates fix commits directly in your existing Git workflow. Findings are validated in simulated environments to cut false positives.
The most mature capability, GA since March 31, 2026. The agent ingests source code, architecture diagrams, and docs to understand how an app was built, then acts like a human tester: identifying vulnerabilities, attempting exploitation with targeted payloads and attack chains, and validating that findings are real risks — not theoretical ones. It connects individual issues into higher-severity attack chains across AWS, Azure, GCP, other clouds, and on-prem.
| Metric | Reported result |
|---|---|
| Timeline | Compressed from weeks to hours in preview feedback |
| Cost | A fraction of the cost of manual penetration testing |
| Coverage | 24/7, on-demand — the full portfolio, not just top-tier apps |
| Findings quality | CVSS scores, severity ratings, reproduction steps, remediation guidance |
Figures are AWS-reported from preview customer feedback — treat as vendor-reported until validated on your own workloads.
AWS positions Security Agent as covering the full lifecycle in one unified offering, with findings flowing between phases:
| Phase | Capability | Typical trigger |
|---|---|---|
| Design time | Design review, threat modeling | New architecture doc, design change, repo connect |
| Development time | Code review (PR + full repo) | Pull request opened, scheduled repo scan |
| Deployment time | On-demand penetration testing | Pre-release validation, scheduled recurring test |
These steps reflect the general onboarding flow AWS documents as of mid-2026. Exact menu names may shift as capabilities move from preview to GA — check the AWS Security Agent User Guide for the current console flow.
Pentesting is GA in six commercial Regions: US East (N. Virginia), US West (Oregon), Europe (Ireland), Europe (Frankfurt), Asia Pacific (Sydney), and Asia Pacific (Tokyo). Design review, threat modeling, and code review are in preview and may have different availability — check AWS Capabilities by Region first.
New customers can start with a 2-month free trial; check the Security Agent pricing page for usage-based costs beyond that.
Connect a repo (GitHub, GitLab, or Bitbucket — SaaS or self-hosted) and, optionally, Confluence for design docs, so the agent reasons from real architecture context.
Select relevant frameworks — AWS WAF, NIST CSF, PCI DSS, AWS best practices — or import internal requirements. Findings map back to compliance posture automatically.
Choose "Enable threat model," connect your repo if you haven't already, and get a STRIDE-based model prioritized by risk.
Configure which repositories are monitored. The agent scans PRs and full repos, returning fix commits and remediation guidance directly in your Git workflow.
Launch a test from the console (or CLI, once integrated) against a target application, and get a report with CVSS scores, severity ratings, reproduction steps, and remediation.
Install the Security Agent power for Kiro, or connect via the open MCP server (Claude Code plugin rolling out separately). Trigger threat models, code reviews, and remediation from your IDE or CLI — e.g. asking to Run a full security scan on this repo.
Use Amazon EventBridge and AWS Step Functions to schedule recurring pentests, with Amazon SNS notifications on completion. AWS publishes a sample CloudFormation template for this exact pattern.
Prioritize by severity, assign fixes, apply suggested remediations or fix commits, and re-run validation before closing anything out.
AWS Security Agent represents a shift from point-in-time, tool-fragmented application security toward a single agentic system that understands an application's design, code, and behavior — and continuously acts on that understanding. For teams facing scaling pressure, the value proposition is simple: security work that used to be periodic, expensive, and expert-gated becomes available on demand, at machine speed, embedded directly into existing workflows.
