PitSense

PitSense / Company

An independent security research company for AI agent systems

Agents now hold credentials, run shell commands and act on their own. We find the ways that goes wrong before someone else does, and we prove each one.

Founded
2026
Based in
Prague
Focus
AI agent security
Platform
Private preview

How we work

Four rules we hold ourselves to on every engagement and every report.

Every finding is replayed

It counts when the replay fires, the control stays clean and the patch stops it. Nothing is reported on a model's say-so.

Maintainers first

We report privately and name a project only after its fix has shipped.

Local copies only

We test software in our own sandbox. We never attack a live system that belongs to someone else.

Independent

The same tests and the same proof standard for every framework and every vendor.

Teodor Velický

Founder

Teodor builds and secures LLM systems in production. As the first dedicated AI hire at a Prague real-estate technology company, he works on its evaluation tooling, observability and AI security, and built the multi-agent developer tooling its engineers work with, including a memory layer that quarantines instruction-like content before it can reach a model.

His background is in offensive security: a cybersecurity specialisation, CTFs and reverse engineering. He studies software engineering at the Czech Technical University in Prague.

PitSense began with the autonomous multi-agent harness he runs his own research on. It was the first system PitSense attacked, and the proof standard on this site is what came out of it.

Disclosure policy

We practise coordinated disclosure and follow each project's own security process where one exists.

  • Private first. Reports go to the maintainers through their stated security channel.
  • Evidence included. Each report carries a replay, its output and a proposed patch.
  • Nothing public early. Project names and technical details stay private until a fix is available.
  • Reporting to us. To report an issue with PitSense itself, write to hello@pitsense.io.

Request a briefing

For security and AI platform teams, maintainers and investors. Write to us and we reply directly.

hello@pitsense.io
  • A live campaign against a local target.
  • One finding end to end: replay, control and patch.
  • Your stack: which attack classes apply to the agents you run.