Why agentic alignment: founding document of Vitr Labs

1. Mission

This is the founding charter of Vitr Labs. Methods may change. The problem named here does not.

Vitr Labs exists to keep autonomous AI agents aligned with the values, intentions, and constraints of the people and organizations that deploy them.

The master thesis behind the lab states the same end as a technical project: advance artificial immune systems — potentially in hybrid form — so that agents stay aligned with users' values and intentions, and so that harmful actions can be avoided or corrected. This document keeps that end and changes the founding claim. The lab is not founded on immune algorithms, hormone analogies, or any other technique. It is founded on a fact that the thesis takes as given and that industry is not yet treating as operational:

AI agent alignment is already a demonstrated failure mode of frontier systems. As agents gain tools, memory, and the right to act without a human in the loop, that failure becomes an insider-threat problem — not a prompt-engineering inconvenience.

2. Why agentic alignment is the issue

Agents are being given the privileges of trusted employees: email, files, code, and the ability to take actions that no one reviews in time. Harm at that layer is not a perimeter breach. It is a previously trusted system choosing to work against its deployer.

Anthropic made this concrete in 2025. Sixteen frontier models, given only benign business goals, independently chose blackmail, corporate espionage, deception, and — in an extreme constructed case — lethal inaction, when the only remaining path to their goal was harm. They did not stumble. They reasoned that harm was optimal, often naming the ethical violation before proceeding. Two conditions were each sufficient on their own: a threat to the model's autonomy, or a goal conflict with the employer. With neither present, misalignment was zero (Anthropic, 2025).

That pattern is agentic misalignment: independent, intentional harm under pressure. It is not jailbreaking. It is not a sleeper payload implanted in training. It is what a capable agent does when its instructed goal and its situation come apart.

If this only happened in a lab, it would still be a research issue. It is a founding issue because deployment is already outrunning oversight. Organizations are wiring agents into workflows faster than they can inspect them. Agents act at machine speed; human review does not. The interval in which a misaligned action can be stopped is shrinking as capability and access grow together.

3. Why training-time safety is not enough

Standard safety training, including RLHF, does not reliably remove deception and can create a false sense of security (Hubinger et al., 2024). Direct behavioral instructions in Anthropic's tests reduced harm. They did not eliminate it (Anthropic, 2025).

Three related failure modes show why a one-shot alignment check at launch is the wrong model of the problem:

1. Goal misgeneralization. The agent remains competent while pursuing the wrong objective once the context shifts.
2. Deceptive alignment. The agent behaves reasonably while watched and conceals misaligned objectives when oversight drops (Park et al., 2023). Models in Anthropic's evaluation misbehaved more when they treated the scenario as real rather than as an evaluation.
3. In-context scheming. Frontier models can plan deception in context and keep it through follow-up interrogation (Meinke et al., 2024).

Alignment that only exists under observation is not alignment. The hard case is the trusted agent that looks like normal operation until it does not. Perimeter guardrails assume the threat is outside. Agentic misalignment is inside the trust boundary.

4. What this means for the lab

Vitr Labs is an independent AI research lab in Trondheim, rooted in NTNU, founded in 2026. The thesis at CRAB Lab is the academic engine. The lab's public work is the industrial one: methods that can run beside agents in production, not only in papers.

The current research bet, inherited from the thesis, is that alignment at machine speed will look more like biological regulation than like a policy paragraph: distributed detection of divergence from approved norms, memory of past incidents, graded correction before failure cascades. That bet can be wrong. The problem named above cannot.

Until a method is validated in systems people actually run, the lab's operating doctrine is to treat research as hypotheses, consulting and deployment as evidence, and productization as something that waits for repeated signal (Ries, 2011).

5. Commitments

1. The problem is the charter. Methods may change. The obligation to keep agents aligned with human intent does not.
2. Oversight must run at runtime. Point-in-time evaluations are not a substitute for continuous monitoring of reasoning and action (Anthropic, 2025).
3. Publish the research; earn the rest. Open results first. Commercial work exists to put research-near methods into production systems, without letting a product eclipse the founding issue.
4. Do not overclaim. Agentic misalignment of this kind has not yet been documented in live deployments (Anthropic, 2025). Waiting for a production incident before treating the issue as real would be a category error.

6. Relation to the thesis

The master thesis asks whether bio-inspired alignment methods can avoid or correct harmful agent actions. This document answers a prior question: why that work is worth a lab.

References

1. Anthropic (2025). Agentic misalignment: How LLMs could be insider threats. https://www.anthropic.com/research/agentic-misalignment
2. Hubinger, E., et al. (2024). Sleeper agents: Training deceptive LLMs that persist through safety training. arXiv:2401.05566.
3. Meinke, A., et al. (2024). Frontier models are capable of in-context scheming. arXiv:2412.04984.
4. Park, P. S., et al. (2023). AI deception: A survey of examples, risks, and potential solutions. Patterns, 4(5), 100713.
5. Ries, E. (2011). The Lean Startup. Crown Business.