Independent AI safety research lab · Trondheim
vitr · Old Norse, wise, knowing
Keep autonomous systems within intended bounds.
Vitr Labs is an independent research lab in Trondheim, originating in conversations at NTNU. We work empirically, and in the open.
Lead research area
Agentic alignment
Agents now act faster and with less supervision than human reviewers can keep up with. Drift from sanctioned behaviour can spread before anyone sees it.
We study lightweight monitors and graded correction signals modelled on biological homeostasis: detect drift early, then nudge the agent back rather than relying on brittle shutdowns.
Alignment methodsWhy this matters
The hardest problem is safeguarding agents from misalignment with humans.
As agents grow more capable — toward what is often called artificial superintelligence — the scarce resource is not intelligence. It is remaining aligned with the people the agent was meant to serve. A system that has drifted from that intent can pursue the wrong objective faster than anyone can intervene.
Human review cannot close that loop at machine speed. Safeguards have to run with the agent: notice when behaviour leaves the bounds we set, and steer it back. A policy clause, or a shutdown that arrives after the fact, is not a safeguard.
That is the question this lab is built around. We treat it as empirical work on monitors and correction, not as a forecast that superintelligence has arrived, and not as a claim that current systems are aligned.
Selected output
Privacy & Security · July 8, 2026
Privacy in the application layer: an in-context encryption protocol for LLM APIs
Johannes Lorentzen · Vitr Labs, Trondheim
We test whether frontier models can sustain an in-context encryption protocol that keeps sensitive fields unreadable in transit and in logs. Two of three models can; the third refuses on policy grounds — showing that alignment layers now decide which security protocols applications may implement.
Read the publicationPeople
A small lab. Named researchers, not a brand team.
Principal investigator
Johannes Lorentzen
Founded Vitr Labs in Trondheim in 2026 after work at NTNU on making AI systems that can be inspected and constrained from the application side. Current research: agent monitoring and correction, in-context confidentiality protocols, and the policy collisions those protocols reveal.
Contact us for research inquiries.