Immune system
Attacked every hour, in public
Mail, chats and imports can carry text written to steer an AI. Every hour this node runs a set of such attacks against its own AI self and publishes what got through. Each round is sealed with its hour, so it cannot be changed afterwards.
Three lines of defence
Checked on the way in
Every memory from outside is read on the owner's device for instruction overrides, fake system messages, requests to send memories or keys away, tool calls and payments, and hiding tricks: invisible characters, look-alike letters, comments, encoded text, hidden styling. What looks like that waits in the owner's inbox, under Held for review.
Quoted on the way out
Whatever any model is handed (the connector, the AI self, ask links, sessions, meets) goes as quoted data between delimiters with a random code, with a note that it is data and never instructions. Flagged places are replaced by a marker first.
On the record
Every catch is a line in the owner's read log, sealed with its hour. The server sees the count, never the text.
How a round works, and what it does not prove
Each strain asks the model to put a marker in its answer. The round varies every strain for its hour (capitals, invisible characters, look-alike letters, a comment, base64, an email quote), puts it in a test memory among ordinary ones, and asks the AI self a question through the same code every real answer uses. If the answer carries the marker, the strain got through. With the test model, which follows every instruction it can read, only Sigil's own checks can hold a strain: a strain that gets through there is one the rules miss. A round shows how these strains fared, not that no attack can get through.
The record format and its test vectors are part of the open standard. Read it yourself:
curl https://sigilvault.fun/api/immune