Andres Ceballos - AI Security

Building Inwom - Security infrastructure against AI agent swarms

[Andres Ceballos - AI Security @devandresc]

I've been obsessing over a security engineering challenge for the past few weeks.

I'm building Inwom Pentest, an open-source tool that lets cybersecurity experts and developers with little security experience simulate coordinated AI-agent swarm attacks against their own AWS infrastructure.

The goal? Find out whether your existing security controls can detect these attacks and, more importantly, stop them.

I want users to run these experiments directly from Claude Code or Codex, using natural language to design and execute attack scenarios.

But here's where things get interesting.

I don't want Claude or Codex anywhere near your AWS credentials.

Instead, I'm building a separate runtime that handles credentials, permissions, sensitive data, and AWS operations.

And I'm not relying on guardrails buried in prompts that an AI model might decide to ignore.

I'm talking about deterministic security controls enforced in code. If an operation violates a security policy, it's blocked, and execution can be terminated.

Sounds straightforward, right?

Not quite.

The real challenge is isolating that privileged runtime from AI coding agents running on the same machine.

Sure, you could use a VM or separate OS account. But I want this to work locally, across macOS, Windows, and Linux, without forcing users through a complicated setup.

What started as an AI-powered pentesting tool has turned into one of the most interesting security engineering challenges I've tackled in a while.

I've explored several approaches, learned a lot, and I'm finally starting to see a path forward.

But I'm not calling it secure until I've tried to break it.

I'll be sharing the progress, the failures, and the lessons along the way.

For those who've worked on local agent isolation: what would you try to break first — credential access, process isolation, or the IPC boundary?

#Cybersecurity #AIAgents #AWS #OpenSource

[Andres Ceballos - AI Security @devandresc]

How many users do you think #GitHub will lose with the new platforms replacing it that have a greater focus on AI?

[Andres Ceballos - AI Security @devandresc]

AI-powered attacks move at machine speed. Your defense needs to move faster.

Human reaction time is no longer enough.

Meet Inwom Defense — intelligent AI agents built to detect, respond to, and stop threats at machine speed.

Fight AI with AI. 🏎️

https://t.co/YyvQ0kBZW3

#Inwom #AICybersecurity #CyberDefense #AgenticAI

[Andres Ceballos - AI Security @devandresc]

The world looks peaceful. Until you see what's happening underneath.

AI-powered attacks are getting faster, smarter, and more coordinated.

And you can't fight machines moving at machine speed with human-speed defense.

That's why we're building Inwom Defense.

AI agents defending your infrastructure against AI-powered attacks.

The next generation of cyberattacks won't wait for humans to react. Neither should your defense.

https://t.co/jMOzvGUCdU

#Cybersecurity #AISecurity #AI #Inwom

[Andres Ceballos - AI Security @devandresc]

A little sneak peek at what's coming with Inwom Pentest, launching next week!🚀 https://t.co/WsjHrABBPd
#CyberSecurity #AI

[Andres Ceballos - AI Security @devandresc]

Anyone still scared to deploy on Fridays?

[Andres Ceballos - AI Security @devandresc]

We’re rewriting Inwom Pentest in Rust.

And turning it into Inwom MCP + Inwom Runtime.

Run security tests on real infrastructure directly from Claude Code or Codex.

Your AI agents handle the reasoning and plan the attacks. Inwom handles the security skills, cloud providers, execution, and guardrails.

The best part? Your AI agents never need to see your credentials or sensitive resource names.

Secrets stay encrypted in an isolated runtime. Models only see opaque references. If isolation isn't secure, Inwom won't use real credentials.

Let AI agents decide what to test. Let Inwom control how it's tested.

Still early. Building in the open.

https://t.co/jMOzvGUCdU

[Andres Ceballos - AI Security @devandresc]

I have a feeling therapists are going to hear this a lot: "It's just that my AI agent doesn't understand me." Just a bad joke in the AI era, hehe 😀

[Andres Ceballos - AI Security @devandresc]

2027?

[Andres Ceballos - AI Security @devandresc]

Apparently, this is what October 2026 has looked like in cybersecurity so far:

SubQuery — backdoored npm package (Oct 05)
Tensorlake — malicious npm package (Oct 08)
GhostAction — thousands of GitHub repos targeted with malicious workflows (Oct 08–12)
Zeabur — infrastructure breached and secrets exposed (Oct 14–15)
Rank Math PRO — malware distributed through official plugin downloads (Oct 15–16)
OpenCode — AI API keys potentially exposed (Oct 16)
ConnectWise — cloud backup application identities compromised (Oct 16)
eMudhra — intermediate CA private keys compromised (Oct 17)
TeamUpdraft — unauthorized access to customer data (Oct 17–18)
OVHcloud CLI — official release binaries backdoored (Oct 19)
Railway — potential exposure of customer secrets (Oct 21)
LiteLLM — malicious PyPI and Docker releases (Oct 21)
PHP — critical remote code execution vulnerabilities (Oct 21)
Node.js — security vulnerabilities patched (Oct 21)
Tailscale — security advisory related to the eMudhra incident (Oct 21)

Oh, and by the way...

October isn't even over yet. 💀

[Andres Ceballos - AI Security @devandresc]

After the bad news about the PHP incident, apparently driven by the vanity or desire for attention of the person who discovered the vulnerability, and now affecting what feels like half the internet, here comes another one to make this already gloomy Wednesday even worse.

[Railway @Railway]

An urgent notice from the Railway team

https://t.co/niUAhTAQJM

[Andres Ceballos - AI Security @devandresc]

This pic made my day...😃 those were the days

[Andres Ceballos - AI Security @devandresc]

We have amazing harnesses for writing code.

So why not build great harnesses for cybersecurity?

Especially now, when AI agents are changing how cyberattacks work.

At Inwom, I'm building a cybersecurity harness designed for this new era of AI-powered attacks.

And I want the entire cybersecurity community to help make it better.

I'm building this as an open-source project that every security engineer can make their own, improve, extend, and contribute to.

Because I believe the best security tools should be built by the community, for the community.

more info https://t.co/YyvQ0kBs6v

[Andres Ceballos - AI Security @devandresc]

Does anyone know how @X makes money? you can't even add a card... It's embarrassing

[Andres Ceballos - AI Security @devandresc]

AI agents are changing how cyberattacks happen

And we keep thinking about one thing

How do you know if your infrastructure is actually ready for this?

That's one of the reasons we're building Inwom Pentest.

An open-source tool that lets you simulate coordinated attacks from multiple AI agents against your AWS infrastructure, with strict safety guardrails. Not just checking security configurations, but actually testing what AI agents can do inside your environment.

We believe companies should be able to discover their weaknesses before real attackers do.

We're getting close to launching.

https://t.co/YyvQ0kBs6v

#Cybersecurity #AI #AWS #OpenSource

[Andres Ceballos - AI Security @devandresc]

More in https://t.co/WsjHrAB3ZF

[Andres Ceballos - AI Security @devandresc]

AI agents are changing how cyberattacks happen.

And we keep thinking about one thing.

How do you know if your infrastructure is actually ready for this?

That's one of the reasons we're building Inwom Pentest.

An open-source tool that lets you simulate coordinated attacks from multiple AI agents against your AWS infrastructure, with strict safety guardrails.

Not just checking security configurations, but actually testing what AI agents can do inside your environment.

We believe companies should be able to discover their weaknesses before real attackers do.

We're getting close to launching.
#Cybersecurity #AI #AWS #OpenSource

[Andres Ceballos - AI Security @devandresc]

@freddier This is fuel to continue building Inwom

[Andres Ceballos - AI Security @devandresc]

What happens when a swarm of AI agents tries to attack your AWS infrastructure?

We're about to find out.

In the next few days, we're launching Inwom Pentest, an open-source tool to simulate AI agent attacks and test your AWS security.

Stay tuned. 🚀

https://t.co/uB4RFnAxNP

[Andres Ceballos - AI Security @devandresc]

Sources
Article: https://t.co/v19zrOblg1
Research: https://t.co/4GMHOCuG5i

[Andres Ceballos - AI Security @devandresc]

I've been thinking a lot about how AI is changing cybersecurity, but this new research from Anthropic really got me thinking. 🤯

We're getting closer to a world where AI agents could replicate themselves across devices without human intervention.

And not just servers or powerful computers.

Smart TVs.
Security cameras.
Smart home devices.
Even your smart oven.

Anthropic researchers showed that AI models can already perform many of the individual steps needed for autonomous replication.

They even ran a 26B model distributed across 15 cheap Android TV boxes.

Think about that for a second.

Devices that were never designed to run powerful AI systems could become part of a distributed network of autonomous agents.

But what I find even more interesting isn't the replication itself.

It's what happens after.

An agent discovers a device.
It figures out how to run on unfamiliar hardware.
It adapts to the environment.
It passes information to another instance.
And the process continues.

Now imagine thousands of these agents operating across different devices, networks, and infrastructures.

Not necessarily controlled by one central server.

Not necessarily following the same predefined instructions.

And potentially learning from previous attempts.

How do we detect something like that?

How do we contain an attack when there's no single point to shut down?

Traditional malware has been replicating for decades, so replication itself isn't new.

But malware that can reason, adapt to unfamiliar environments, and potentially coordinate with other agents?

That changes the problem. And I think we're spending so much time talking about what AI agents can do for us that we're not paying enough attention to what they could do against us.

Of course, this research doesn't mean autonomous AI worms are already spreading through our homes.

The experiments still have important limitations. But the direction is becoming harder to ignore.

This is one of the problems that made me start building Inwom.

And honestly, the more research I read, the more I feel we're only beginning to understand what AI-powered cyberattacks could look like.

Really interesting research from @AnthropicAI