Hugo Pereira

Cybersecurity since 2014. AI psychosis since 2026. Fortinet Certified Trainer. Hermes Agent contributor. Least privilege for users, root for my agents.

[Hugo Pereira @HugoPereira84]

Least privilege for the users, root for my agents.

I'm Hugo, a Brazilian living in Norway. Networks since 2001, security since 2014, with work in Brazil, Africa, Australia and Norway. Fortinet Certified Trainer and 2x FCSS (Enterprise, OT) at Exclusive Networks.

On the side I build AI agents that do real IT and security work. Here I'll post what works, what breaks, and what I've learned from giving them root.

My opinions are my own and not those of my employer.

[Hugo Pereira @HugoPereira84]

@witcheer Thanks! Nice to finally contribute something back to a project I use so much

[Hugo Pereira @HugoPereira84]

2 decades of telling customers to patch faster, and my first open-source contribution is making an AI agent wait less. Priorities.

If you use Hermes Agent daily, you know the pause while it compresses its context. Now it can write that summary in the background before it's needed, so you just keep working.

It's off by default. To turn it on:
hermes update
hermes config set compression.prepare_ahead true

Thanks @Teknium and Kshitij for taking it the rest of the way.

[Hugo Pereira @HugoPereira84]

@lukensort Would love to hear more about how you've set that up.

[Hugo Pereira @HugoPereira84]

@SmashGoCo Hot take: the cyberpocalypse hasn't even started. I suspect a huge number of environments have obvious holes. We're only being forced to face it now because the cost of hacking is falling through the floor.

[Hugo Pereira @HugoPereira84]

@theo Finally taking the plunge. Late to the party 😉

[Hugo Pereira @HugoPereira84]

Running Tailscale SSH? Until there's a patch for Tailworm: write SSH rules as one device to one target, turn off SSH forwarding, enable node-key expiry.

On headscale you can't turn forwarding off, so treat every SSH grant as root. https://t.co/CMaIUoqdBu

[The Hacker News @TheHackersNews]

‼️ WARNING - A researcher just dropped a report on Tailscale SSH flaws — one compromised device could get root on others, and there’s NO PATCH yet.

In lab tests, SSH port forwarding enabled privilege escalation and lateral movement across already-permitted devices.

Tailscale says a fix is coming.

Read - https://t.co/0wcJTvDoxV

[Hugo Pereira @HugoPereira84]

@secdetect I’d like to see more scrutiny of the MCP servers themselves. We’re still installing someone else’s code and giving it access.

[Hugo Pereira @HugoPereira84]

@HermesWatcher Probably a skill with a script behind it. I’d add an MCP server if it saved me maintaining the integration myself.

[Hugo Pereira @HugoPereira84]

@HouseofChimera I’d really like to know what those two audits checked. How does something like this go unnoticed?

[Hugo Pereira @HugoPereira84]

@TheHackersNews I’d be checking the SSH rules here. How many other machines could a compromised laptop actually reach?

[Hugo Pereira @HugoPereira84]

@IntCyberDigest Authenticator already knew which app was calling. It just didn’t use that information to check the request. That’s a frustrating one.

[Hugo Pereira @HugoPereira84]

eMudhra's intermediate CA private keys were 4, 56, 66, 115, 155, 156, 230 and 237. Made on certified HSMs, passed two audits, live for 18 months. Mozilla says local AI agents cracked them on the first try with a CTF prompt. https://t.co/Ohiv5Mjj4C

[stacksmashing @ghidraninja]

So the private-key for the eMudhra (revoked) intermediate was apparently 4.

How is the root still trusted? How broken are your processes if this can happen?

And honestly: How have we (as an industry) not noticed earlier :(

[Hugo Pereira @HugoPereira84]

Also worth knowing if you run long Hermes sessions: when the context fills up, Hermes stops to compress it, and the median pause was about 3 minutes. A new opt-in setting writes that summary in the background before it's needed, so the agent keeps working instead of freezing mid-task.

hermes update
hermes config set compression.prepare_ahead true

Details 👇 https://t.co/lZxMsn3YkZ

[Hugo Pereira @HugoPereira84]

2 decades of telling customers to patch faster, and my first open-source contribution is making an AI agent wait less. Priorities.

If you use Hermes Agent daily, you know the pause while it compresses its context. Now it can write that summary in the background before it's needed, so you just keep working.

It's off by default. To turn it on:
hermes update
hermes config set compression.prepare_ahead true

Thanks @Teknium and Kshitij for taking it the rest of the way.

[Hugo Pereira @HugoPereira84]

@theo I feel like I can discuss ideas with Fable, not how to do something, but if something was the right thing to do. I could get 1 layer of abstraction higher with Fable

[Hugo Pereira @HugoPereira84]

@theo I deeply believe in agents having memory, but I don’t think it should be enterprise-wide, but team-wide. My infra agent doesn’t need to know what the accounting team is up to, but I cannot overstate the helpfulness of having a “multiplayer” agent my team also feed with context

[Hugo Pereira @HugoPereira84]

@theo Anybody else preferring Opus 4.6/7/8 to 5?