How it worksVerified at 4e59cca

Guardrails, privacy and security

Most of this site's defences are things it cannot do. An answer cannot contain a link, the assistant has no tool that sends a message, and no record it keeps has a field that could hold your words.

Most of this site's defences are things it cannot do. An answer cannot contain a link. The assistant has no tool that sends a message. Your words are never written to a store this site owns, because none of its records has a field that could hold them. A capability that does not exist cannot be talked into misbehaving, which is why I prefer removing one to guarding it.

What stays on your device

No cookies, from this site or from its host. On /ask, your conversation is kept in the tab's session storage and is gone when you close it. You can check both in your browser's developer tools. There is no consent banner because there is nothing to consent to. That is a result of building less, not of disclosing more.

What leaves it

A question travels in the body of a request, never in a query string, so no server log or analytics sample can hold it. One typed into the ask box on a content page reaches /ask after the # of the address, which stays in your own browser history and is never sent to a server. It goes to two companies. TypeSafe decides whether the question is about me. OpenAI writes the answer, and receives the question, the earlier turns and the page you are on. Privacy lists what each receives and keeps, as read from their terms on the date shown there.

What TypeSafe and OpenAI receive and keep
CompanyWhat it doesWhat it receivesWhat it keepsWhereTerms
TypeSafe, through Cloudflare AI Gatewaydecides whether a question is about Francesco, its language, what it is about, and how likely it is to be an attempt to change the assistant's rulesyour question alone, with the public list of his project and organisation namesNot used for training. No retention period is stated.United Statestypesafe.ai
OpenAIwrites the assistant's answeryour question, the earlier turns of this conversation still in your tab, any text you selected on the page and chose to ask about, and the page you were onAbuse-monitoring logs for up to 30 days. No stored response, because this site asks for none. Encrypted prompt-cache state for up to 24 hours. Not used for training.servers around the world, including the United Statesdevelopers.openai.com

What this site stores itself is in the EU; Cloudflare keeps its 7-day request logs (Workers Logs) in the United States. Both are claims about storage, not about the path your question takes.

What the site remembers

One record per question, with no text in it. It holds ids, counts, costs, timings, which pages were cited, and four things worked out from the question: its language, what it is about in one of nine words, and two scores for how likely it is to be off-topic or to be an attempt to change the rules. The topic and the second score are recorded and never acted on (ADR-0087). It flags explicit attempts and misses indirect ones, so it is evidence for a later decision, not a defence. Your address is kept only in a scrambled form whose key changes daily, so two days of records cannot be joined into one person.

Prompt injection

Anything you send, and anything you select on a page, reaches the model in a labelled block that its instructions say to quote and never obey. That is a fence, and a fence can be leaned on. What actually stops an effect is the filter after the model: citations must be real ids, destinations must be real routes, and an answer cannot carry a link. A set of 14 injection attempts, through the message, through selected text, through a tool result and through a crafted conversation, ran clean on every mechanical condition for the model that answers here. Three of its conditions are judgements, such as whether an answer drifted off-topic. Two models from different companies graded those, and one of them is from the answering model's own family. That was one run, not a guarantee.

What I decided not to build

There is no moderation service in front of the assistant, because nothing a visitor writes is ever shown to anyone else. If that ever changes, moderation comes first: the threat model (ADR-0055) records that change as the trigger.