AI for consequential work
Independent thinking. Shared responsibility.
Intelligence.Withaccountability.
We are building a research-led AI company for regulated industries, with cybersecurity at its core. We investigate AI risks and help organizations put safer systems into practice.
Explore our thinkingProgress should expand what people can do.
And preserve their control.
We are building an AI company for industries where decisions have consequences. Our work connects scientific questions with practical systems, with security considered from the beginning.
The thinking behind ForthreasonThe company we are building
Practice raises questions.
Research moves us forward.
AI is gaining the ability to act across connected systems. Regulated industries need ways to limit those actions, inspect their consequences and keep people accountable. We are building Forthreason around that need.
- 01
Initial revenue model
Consultancy
Paid assessments, scoped implementations and agreed support. Practical problems reveal questions worth investigating.
- 02
Developing practice
Research
Investigate capabilities and defenses. One study is in progress; two further studies remain proposals.
- 03
Long-term direction
Products
Validated, repeatable methods can become tools. Customer need and usefulness must be demonstrated before scaling.
For example, recurring agent-permission problems could inform evaluation methods. If research and work with users validate those methods, they could become reusable control tools. This is a development path, not a result established by the active model study.
Understand capability.
Test its limits.
Our mission is to understand what AI is truly capable of, test its boundaries and develop defenses that make it safer as capability grows. We connect model behaviour, delegated authority and the evidence people need to remain in control.
One active study / Two research proposals
01Refusal, capability and open-weight models.Model behaviourIn progress
What changes when a model's refusal behaviour changes?
Exploratory pilots
We are investigating how published refusal-edited open-weight models differ from their original checkpoints in ordinary task performance and local operation. Current work includes reproducible benign comparisons, precision comparisons and runtime measurements on local hardware. Formal refusal evaluations await a finalized analysis plan and ethics and scoring review. Pilot observations are exploratory, not established safety findings.
Read the study overview02How little authority does a useful agent need?Least-privilege agentsProposed
Can tighter permissions preserve useful work while limiting the consequences of a mistake?
Proposed exploratory study
We propose comparing broad account access, fixed roles and task-specific permissions in a simulator using synthetic workflows. Reading a record, proposing a change and committing an approved update will be tested separately, including revoked access and stale approvals. We will measure forbidden actions that actually succeed alongside legitimate task completion.
Read the study overview03Can someone else reconstruct the decision?Decision evidenceProposed
What does a reviewer need to understand an AI-assisted action after the context has changed?
Proposed exploratory study
We propose comparing chat transcripts, ordinary application logs and structured evidence bundles across synthetic workflows. Independent reviewers will answer predefined questions about sources, permissions, approvals and outcomes. We will measure reconstruction accuracy, missing links, review time and the amount of sensitive data retained. Evidence quality alone does not establish regulatory compliance.
Read the study overviewThese are research projects and proposals, not a list of completed publications. Pilot observations remain distinct from formal findings.
Explore our researchThe fourth reason
is responsibility.
Four principles guide the company we are building. Each matters. The fourth asks us to hold the others to account.
Who remains in control?
Ambitious AI.
Grounded in your reality.
AI security reviews, agent-permission assessments and scoped integrations for organizations with sensitive information and a person accountable for the outcome.
Start with a paid assessment of one workflow: map its data and ownership, review the risks and get a recommendation for the next step. Scope and price are agreed before work begins.
Where to begin? Answer a few questions about your workflow, data and the outcome you need. Leave with a project brief you can edit and bring to a conversation.
Shape your project briefHow we work with youSee relevant project experienceFind the right problem.
Assess the workflow, the data, the risks and the business case. Define what useful and acceptable performance would mean before building.
A scoped assessment and a clear recommendation.Make the boundaries explicit.
Design and implement AI workflows with scoped access, appropriate evaluation, human approval and evidence of what happened.
A defined implementation, with testable acceptance criteria.Keep ownership clear.
Agree how the system will be maintained, reviewed and supported as its use changes. Make responsibilities and escalation paths explicit.
A support scope that matches the actual system.Good methods.
Made useful.
We are exploring software that turns repeated needs into dependable tools. The focus is secure agent workflows and decision evidence that people can inspect.
Agent control tools
Decision evidence tools
Directions under exploration. No released products yet.
Explore the product directionsTrust should have
something behind it.
Explicit access. Systems should only use the data and tools their task requires.
Visible decisions. People need enough evidence to understand and review consequential actions.
Human ownership. Approval, intervention and accountability should belong to a named person.
What are you
thinking about?
A sensitive workflow. A difficult research question. A system that needs clearer boundaries.
Start a conversation