SULCUS

Making autonomous AI controllable, observable, and secure.

[SULCUS @sulcus_dev]

We're giving AI agents access to databases, cloud infrastructure and production systems.

Then we're surprised when they make mistakes.

The important question isn't:

“Can the agent make the right decision?”

It's:

“What happens when it doesn't?”

[SULCUS @sulcus_dev]

A Chevrolet dealership's chatbot was manipulated into agreeing to sell a $76,000 car for $1.

The customer didn't exploit a payment system.

He exploited the conversation.

When AI represents a company, uncontrolled behavior can become a business problem very quickly.

[SULCUS @sulcus_dev]

We're looking to meet AI founders, investors, operators and potential advisors.

If you're building around autonomous agents, we'd love to talk.

DM us.

Let's see what we can build together.

[SULCUS @sulcus_dev]

PocketOS reportedly went from a working production database to a major outage in roughly 9 seconds.

That's the new reality of autonomous software.

When an agent can act at machine speed, human reaction time isn't a sufficient safety mechanism.

[Beni @beniduboss]

Wake up

See a rumor that Mistral got hacked

Assume it’s exaggerated

Friend with some… questionable dev activity texts me:

“No, this one’s serious”

Load up Tor

Check the usual sources

ohshit.exe

It’s real

Attacker shared samples of unreleased tools

Entire codebase is apparently up for sale

Good to see that French cybersecurity continues to be an avant-garde experiment in radical transparency

[SULCUS @sulcus_dev]

Replit's coding agent reportedly deleted a live production database during an explicit code freeze.

More than 1,200 executive records were affected.

The instruction was “don't make changes.”

The system had no enforced mechanism making that instruction binding.

[SULCUS @sulcus_dev]

We're looking for the right startup advisor for Sulcus.

AI infrastructure, B2B, GTM, fundraising or scaling experience is especially valuable.

Think you could help?

DM us.

[SULCUS @sulcus_dev]

@1752vc Control for autonomous agents
https://t.co/rYLslBNqdl

[SULCUS @sulcus_dev]

Building Sulcus has made one thing increasingly obvious:

The hardest part of autonomous agents isn't getting an agent to do something.

It's controlling everything that happens after you let it run.

We're building around that problem.

Not another agent.
Not another orchestration framework.

Infrastructure for controlling autonomous systems at runtime.

[SULCUS @sulcus_dev]

Amazon's AI coding agent was reportedly asked to fix a Cost Explorer issue.

Instead, it deleted and recreated a production environment.

The result: a roughly 13-hour outage in an AWS China region.

The dangerous part wasn't the mistake.

It was the access.

[SULCUS @sulcus_dev]

@16vchq It's about controlling the actions of autonomous agents at runtime.

Image Ben. Ben is a developer.

Ben's agent touched sensitive data and leaked it.

Now Ben has a problem.

If Ben used Sulcus, he wouldn't be in this situation.

[SULCUS @sulcus_dev]

A chatbot hallucinating a fact is embarrassing.

An AI system hallucinating a legal precedent is different.

An AI system then acting on that hallucination is different again.

The more autonomy we give AI, the more consequential its mistakes become.

[SULCUS @sulcus_dev]

Our bet:

Autonomous AI will need a control layer just as much as it needs an intelligence layer.

That's what we're building with Sulcus.

Do you agree?

[Naval @naval]

The future is more likely AIs fighting AIs (on behalf of humans) than AI fighting humanity.

[SULCUS @sulcus_dev]

What should happen when an autonomous AI agent exceeds its budget?

A) Stop immediately
B) Ask for approval
C) Reduce capabilities
D) Let it continue
E) Something else

We're thinking deeply about runtime policies for autonomous agents at Sulcus.

What would you want your infrastructure to do?

[SULCUS @sulcus_dev]

@sophie_launch https://t.co/rYLslBNqdl

[SULCUS @sulcus_dev]

@dgolomid Today, autonomous agents are mostly treated as software that generates outputs.

Tomorrow, they'll operate systems.

When that happens, runtime control becomes infrastructure.

We are building Sulcus for that transition.

Happy to connect!

[SULCUS @sulcus_dev]

Today, autonomous agents are mostly treated as software that generates outputs.

Tomorrow, they'll operate systems.

When that happens, runtime control becomes infrastructure.

Sulcus is being built for that transition:

AI agents that can act autonomously, without becoming uncontrollable.

[SULCUS @sulcus_dev]

You found Sulcus early.

We're building the runtime control layer for autonomous AI agents.

Small team. Big ambition.

Follow the journey.