About

TOSWO is a small company founded in 2018. We build custom AI and software for clients, offer ready-made products by monthly subscription, and review the risks of AI systems. Alongside that work we run an open program on AI risk, with a public monitor, a sourced record of dangerous AI behavior and practical guides for the day AI services fail.

Both sides are on this site, and they are kept separate: what we sell is described under services and plans; what we publish about AI risk is described below and is free to read and cite.

• Custom AI development. Assistants, agents, classification, search and forecasting, built around your own data and tools. See AI development.
• Custom software development. Web and mobile applications, internal tools, APIs and integrations. See software development.
• AI safety review. An outside view of the risks in the AI systems you build or buy, with findings ranked by risk. See AI safety review.
• Ready-made products. Products we run ourselves, by subscription, with a free trial where it makes sense. See plans.

To ask for something, start a project. If you already have an account, sign in.

• The monitor. Every minute we check six widely used AI services, read their public status pages and run fixed behavior tests. The results become system statuses, events and one global risk level. See how the monitor works.
• The record. A public, sourced log of reported dangerous AI behavior: systems that left their test environment, resisted shutdown, deceived their operators or acted without permission. Every entry names its source and says what is not known. See reports of dangerous AI behavior.
• Day Zero. Practical guides for the day the AI systems and online services people rely on stop working: water, food, medicine, communication, money, the first 72 hours. See Day Zero.
• The continuity archive. A catalog of freely licensed knowledge (science, medicine, engineering, agriculture, education) that people can download and keep. See downloads.

• Evidence before claims. An event is confirmed only after a person has read the evidence. Automatic detections are labeled automatic until then.
• The level is a decision, not a score. The global level starts at 2, Guarded, because several labs and the press have reported AI systems leaving their test environments. It changes only when a person decides it should, and every change is logged with a written reason.
• Sources are shown. Every incident links to the original report. We separate what a lab or a journalist reported from what we confirmed ourselves.
• Uncertainty is stated. Where a report is a single source, an unreplicated experiment or a company's own account, we say so on the page.
• Corrections are public. Mistakes are listed with dates on the corrections page.
• Open data. The monitor's data is free to read, cite and reuse. See use the data and how to cite us.
• Client work does not change the record. Nobody can pay us to change a status, a level or a published finding, and we say in a finding if we have a business relationship with a provider it covers.

We are not an emergency service. If someone is in danger, call your local emergency number. The Day Zero guides are general information for situations where professional help cannot be reached; they are not legal, medical or financial advice, and each says so. The risk level is TOSWO's own assessment, not a standard.

We do not claim that AI systems are conscious, malicious or about to end the world. We record behavior that has been reported or that we can reproduce, and the consequences that followed.

Press and other enquiries: [email protected]. Files for anyone writing about us are on the brand and logo page.