AWS Security Kit · Agentic AI
Secure your agentic AI workloads on AWS
The AWS Security Hardening Kit for Agentic AI is a practitioner kit for locking down real AI systems running on AWS — voice agents and agentic workflows on Lambda, API Gateway and DynamoDB. It is not a generic AWS checklist: alongside the infrastructure baseline it closes the gaps that matter for an AI stack — third-party AI provider keys in Secrets Manager, least-privilege IAM per agent, network isolation, and containing the blast radius of an agent that can act on its own — the things CIS and the AWS benchmarks don’t reach. You get a 111-control audit checklist across 8 domains, a hardening playbook, six ready-to-apply IAM policies, and a threat model. Instant download, one file you own.
25+ years cloud & AI production experience · PMP · AI Practitioner
Every control in this kit reflects a real incident or near-miss from production AI deployments — not a theoretical framework. If you’ve shipped an AI workload and know it needs hardening before go-live, this is the checklist you were going to have to write yourself.
Get the kit Instant download · No subscription · 14-day refund if it isn’t useful“Why not just use CIS or the AWS benchmarks?”
Because none of the standards, on their own, tell an AI builder what to actually do on AWS. General benchmarks secure the account but ignore the AI layer; AI frameworks name the AI risks but aren’t AWS-specific. This kit is the intersection.
Secure the account — but AI-blind
- CIS AWS Foundations Benchmark
- AWS Foundational Security Best Practices
- AWS Well-Architected — Security Pillar
- NIST 800-53 / CSF, ISO 27001, SOC 2
Secure a generic account. They don’t tell you how to isolate a third-party AI key, scope IAM to a single agent Lambda, or lock down a voice-agent webhook.
Name the AI risks — but cloud-agnostic
- OWASP Top 10 for LLM Applications (2025)
- OWASP Top 10 for Agentic AI (2025)
- NIST AI RMF · MITRE ATLAS
Tell you what to worry about, not the concrete AWS IAM policy, secrets pattern or network control to ship.
The AI threat model → real AWS controls
- Maps the AI-workload threat model to specific AWS controls
- 111 controls across 8 domains, severity-rated
- Six ready-to-apply IAM & resource policies (JSON)
- Secrets isolation, least-privilege IAM & blast-radius containment for the agent stack
AWS-specific and AI-specific — and drawn from real production incidents.
What’s in the kit?
More than a checklist. Five components that take you from audit to fixed — with the policies already written.
111-Control Audit Checklist (Excel)
Every control across 8 domains, each severity-rated (Critical / High / Medium / Low) with guidance and an evidence field, plus a built-in remediation tracker.
Security Hardening Playbook (PDF)
The narrative companion: why each domain matters and how to implement each control correctly, drawn from hands-on production AI deployments.
6 IAM & Resource-Policy Templates (JSON)
Ready-to-apply, least-privilege policies: Lambda voice-agent execution role, S3, DynamoDB, Secrets Manager, cross-account trust, and CloudTrail bucket policy.
Threat-Model Diagram
The agentic AI attack surface on AWS in one picture — the data flows, trust boundaries and where each control sits.
Setup Walkthrough (PDF)
A guided path through applying the kit to a real serverless AI stack, greenfield or existing.
Severity + SLA framework
Critical (fix in 24h), High (1 week), Medium, Low — so you fix what’s dangerous before what’s cosmetic.
Built around a real voice-agent stack
The kit was written and pressure-tested on a production AI voice-agent architecture on AWS — a caller hits an API Gateway webhook, a Lambda function orchestrates the conversation, third-party voice APIs (ElevenLabs, VAPI, Bolna, Plivo) do speech, keys live in Secrets Manager, and state persists in DynamoDB. Every control, IAM policy and threat-model boundary maps to this exact flow, so you’re copying from something that actually shipped — not a generic diagram.
The stack is the example, not the requirement — the same 111 controls apply to any agentic or serverless AI workload on AWS.
111 controls across 8 domains
The infrastructure baseline every AWS account needs — plus a full domain dedicated to the AI-specific risks the standards miss.
Secrets & Credential Management
Every third-party AI key (ElevenLabs, VAPI, Bolna, Plivo, OpenAI) in Secrets Manager with rotation — never in Lambda env vars or a committed .env.
IAM & Least Privilege
One Lambda = one scoped IAM role. Kill the Action:"*" Resource:"*" blast radius that turns one compromised function into full account takeover.
Network Security
VPC endpoints, private subnets, and CloudFront secret-header patterns for serverless AI — keep Secrets Manager and model traffic off the public internet.
Monitoring, Logging & Alerting
CloudTrail integrity, GuardDuty, and immediate alerts on DeleteTrail / StopLogging — an attacker covering tracks.
Data Encryption
Correctly-configured KMS/CMK encryption at rest and in transit — because “encrypted by default” is not the same as configured right.
AI-Specific Security Risks
Securing the AI layer itself: third-party provider keys isolated in Secrets Manager, model version pinning, and containing what an autonomous agent is allowed to do — so a compromised agent can’t reach beyond its own scope.
Compliance & Governance
Coverage that extends past AWS to the AI providers in your data flow, the models used, and the data they process.
Incident Response
Contain-and-recover runbook for AI workloads — disable a compromised key or Lambda URL fast, then rotate and review.
Who is it for?
The people on the hook when an AI system goes to production.
AI agency operators
You deploy voice agents (ElevenLabs, VAPI, Bolna, Plivo) on AWS for clients and need to prove the stack is secured before go-live.
Cloud architects & DevOps
You build agentic AI pipelines on Lambda, API Gateway and DynamoDB and need least-privilege IAM that’s actually correct.
CTOs & technical leads
You’re running a security review of your AI stack and need a defensible, evidence-backed baseline — fast.
Frequently asked questions
Why do I need this if I already follow CIS or the AWS benchmarks?
CIS AWS Foundations, AWS FSBP and Well-Architected secure the account and infrastructure baseline — but they’re written for generic workloads and don’t tell you how to isolate a third-party AI provider key, scope least-privilege IAM to a single agent Lambda, or contain the blast radius of a voice-agent stack. AI risk frameworks (OWASP LLM Top 10, OWASP Agentic AI, NIST AI RMF, MITRE ATLAS) name the risks but are cloud-agnostic and don’t give you the concrete AWS control. This kit is the intersection: the AI-workload threat model turned into 111 specific AWS controls and ready-to-apply IAM policies.
What exactly is in the kit?
A 111-control audit checklist (Excel, 8 domains, with a remediation tracker), a security hardening playbook (PDF), six ready-to-apply IAM & resource-policy templates (JSON), a threat-model diagram, and a setup walkthrough (PDF).
What’s AI-specific about it?
Secure handling of third-party AI provider keys (ElevenLabs, VAPI, Bolna, Plivo, OpenAI) in Secrets Manager rather than Lambda env vars, model version pinning, least-privilege IAM for serverless AI on Lambda / API Gateway / DynamoDB, and containing the blast radius of an agent that can take autonomous actions.
Do I need a specific AI provider for it to apply?
No — it works with any AI stack on AWS. The controls apply whether your agent calls OpenAI, a hosted model, or a third-party voice platform like ElevenLabs or VAPI. The kit covers the AWS side (IAM, secrets, network, logging, encryption) and how your keys and data flow to whichever provider you use.
What’s the refund policy, and how is it delivered?
Instant download after checkout, no login required. If it isn’t useful, email within 14 days for a full refund — no forms, no friction.
Harden your AI stack before it ships
Audit it against 111 controls, apply the IAM policies that are already written, and walk into your go-live review with evidence — not hope.
Get the kit Instant access · No subscription · 14-day refund if it isn’t useful