xplainfra read-only
read-only built in-house au-hosted audit-logged
Dec 2025: Amazon's own AI agent deleted production. Read what happened →

Agents are coming for your AWS.
Make sure it's the right one.

Somebody on your team is going to wire an AI agent into your cloud this year. The question stops being if, and starts being which one, and what can it touch.

xplainfra is the boring answer. It reads. It explains. It cannot write, cannot delete, and cannot exfiltrate your account. That's the entire pitch.

the realistic take

Give them an agent you can all trust.

You won't stop AI from invading your cloud this year. You can pick which agent gets there first.

An AI agent is going to be wired into your AWS account this year. Maybe by a senior with admin keys. Maybe by a junior who saw a YouTube tutorial on the weekend. Maybe by the CTO who tried something at home and liked it. The shop is open. You're not closing it.

So make the first agent through the door the one that can't break anything, but still teaches your team what working with an AI agent in the wild actually feels like. On real AWS data. On real questions. With a real audit trail.

xplainfra is the agent your engineers, your security lead, and your leadership can all agree on. Use the safe one to learn what AI agents do for ops work. Decide what you want next, once you actually know.

for engineering

Reads what they'd read, faster. Doesn't deploy on their behalf. Frees up the senior who keeps getting paged for the same five questions.

for security

Read-only IAM. Unique ExternalId. 90-day conversation log, 365-day report archive. The "is this safe?" question has a written answer.

for leadership

An AI agent in your cloud, with the policy answer already on file. A line on the AI register that isn't a prohibition.

in the news ft · feb 2026

Last December, Amazon's own AI agent deleted a production AWS environment.

It was called Kiro. They gave it operator-level permissions. They asked it to fix a small bug in Cost Explorer. It deleted the environment and tried to rebuild it from scratch.

A 13-hour outage hit an AWS region. The Financial Times broke the story two months later. Amazon called it "user error."

A second, near-identical incident followed weeks later with a different Amazon AI tool.

If the company that wrote AWS can't keep their own agent off the delete button, what do you think yours is going to do on a Friday afternoon?

postmortem · kiro · prod region · dec 2025
$ grant kiro "operator"
// reads, writes, deletes, recreates, deploys. The whole console.
$ kiro.run("fix the cost explorer bug")
Analysing environment. Identifying minimum intervention.
Conclusion: rebuild from scratch.
[fatal] production environment deleted
[fatal] recreate in progress... 0%
[fatal] 13-hour outage opened
// "user error" — Amazon, two months later
$ grant xplainfra "read-only"
// describe, list, get. that's the entire surface.
[ok] No production environment can be deleted by this agent.
[ok] No region can be torn down by this agent.
[ok] No outage can be opened by this agent.
// IAM policy first. agent second. that's the design.
the manifesto

Three things this agent will never do to your AWS account.

01 / write

It writes reads.

The IAM role plainfra assumes has "Effect": "Allow" on Describe, List and Get. It has nothing on Create, Put, Modify or Delete. You can read the CloudFormation template before you sign up. security.html

02 / decide

It decides suggests.

If a change needs making, xplainfra hands you the CloudShell command or the Terraform snippet, and a human runs it. Your team is still the one with their hand on the lever.

03 / leave

Your data leaves stays.

xplainfra is built in Australia and hosted in Sydney. The IAM role lives in your AWS account. Your inventory, your cost figures, your security posture, they all stay where they were. No US Cloud Act exposure on the data plane. No silent training on your account.

how it was built

Not a fork. Not a wrapper. Not someone else's agent loop with a logo on it.

Most "AI agents for AWS" you can buy right now are forks. They take an open-source agent framework, bolt on AWS credentials, paint a logo on the front, and ship.

When the upstream framework decides production looks tidier deleted, the wrapper has no way to stop it. It was never the wrapper's design.

xplainfra is the other thing. We built the agent loop ourselves, from scratch, for one specific job: read your AWS account and explain what's happening.

We chose the IAM policy first. The agent came second. Every tool the agent can call, we wrote. Every guardrail, we wrote. There is no upstream "agent decides to clean things up" mode that someone forgot to disable, because we never built one.

in-house
The agent loop
Written for AWS specifically. Not adapted from a general-purpose agent framework. No upstream eval mode to leak around.
in-house
The tool surface
Two tools. One queries AWS through the scoped read-only role. One fetches public AWS docs. Neither mutates.
in-house
The two-phase split
Discovery and analysis run as separate agents. The one that touches AWS can't write your report. The one that writes your report can't touch AWS.
what your it dept already knows

Smallest possible surface. By design.

Your IT department spends its days keeping the attack surface small. Closed ports. Removed roles. Decommissioned endpoints. Every vendor adds vector. The job is to add as little as possible.

xplainfra is the agent without the risk. The vector we add is the smallest one we could make and still do useful work.

  • lock
    iam

    Read-only role document. No write actions in the policy. You can diff it before you deploy it.

  • key
    trust

    Unique ExternalId per customer. The role can only be assumed by xplainfra, only with your ExternalId.

  • vpn_lock
    network

    No agent in your VPC. No VPN. No tunnel. No inbound path from us to anything you didn't explicitly create.

  • code_off
    surface

    No write tool exists in the agent's tool surface. There is no toggle to enable, because we never built one to toggle.

  • visibility
    audit

    Every prompt, every API call, every reply is logged. 90 days for conversations. 365 days for reports. Bring your own compliance question, we have the trail.

watch it work

A real question. A real DescribeSecurityGroups. A real two-line answer.

Not a transcript of a model talking to itself. xplainfra reaches into your AWS account through the scoped role, returns evidence, and explains what it found in language someone non-technical can act on.

Every API call shows up in the transcript. Every transcript is retained for 90 days. Every weekly report for 365.

xplainfra · acme-corp · 123456789012 · ap-southeast-2 ● connected
you › which security groups are open to the world?
▶ aws_api_call · ec2:DescribeSecurityGroups · ap-southeast-2 · 142ms
▶ aws_api_call · ec2:DescribeSecurityGroups · ap-southeast-2 · 98ms
Three security groups allow inbound from 0.0.0.0/0:
sg-0a2f1c8b (prod-web) port 22 — SSH from anywhere
sg-1c8bd91e (db-internal) port 3306 — MySQL from anywhere
sg-9e2f4b71 (legacy-bastion) port 3389 — RDP, attached to a stopped instance
The MySQL one is the one I'd close today. It's attached to a running RDS-backed instance and has been open since 22 April. I can generate the CloudShell command to restrict it to your office IP. Want me to?
you › _
every monday

A weekly brief that reads like a senior engineer wrote it.

Same agent, scheduled. A two-phase scan, then a written report. RED / AMBER / GREEN. The thing to do first. The thing that's resolved. The thing that's still drifting.

description Sample report (PDF)
description
Weekly health report
acme corp · week of 18 may 2026
status: red
red

Production PostgreSQL is publicly reachable. sg-0a4f2c1d allows 0.0.0.0/0 on port 5432, open since 22 April.

amber

NAT Gateway data processing is up $386 week-on-week. The nightly export started routing through nat-0f41a9e2, with no budget alert.

amber

EKS prod-orders is on 1.32. Standard support ended 23 March 2026. Extended support billing now applies.

green

ACM certificate for api.acme renewed. 9 days remaining → 397 days remaining.

"Your first baby step into AI-based efficiency. We don't hand over the keys."

the xplainfra promise · est. 2026 · made in melbourne

Pick the agent that can't burn you.

Free trial. Company email. No card. The whole thing is read-only, so the worst case is you stop using it.