Home / Security and data

What leaves the machine, what gets archived, what never gets shared

A request sends the prompt text and the model's responses over the network, encrypted in transit, and the local copy of that exchange then stays readable in plain text on disk for a duration fixed by default; deleting a file leaves no way back, whereas archiving it keeps it available and reversible; and an internal document rich in detail becomes dangerous the moment it leaves, for the very reason that makes it useful internally, which calls for a raw inventory followed by an adversarial re-read before any sending.

A conversation with Claude Code sends the text of the request and the model's responses over the network, encrypted in transit by TLS 1.2 or a later version, according to the product's official documentation. What then remains on the local machine is not erased for all that: session transcripts are kept in plain text on disk, in the user's projects folder, for thirty days by default, to allow an interrupted session to be resumed. One documented exception sets this limit aside: a session started or last resumed in Claude Desktop or Cowork is by default exempt from this thirty-day limit, so its transcript stays in plain text on disk with no automatic deletion scheduled.

What a request actually sends

Two facts complete each other. The first: every prompt and every model response leaves the machine the moment a request is sent. The second, less visible: the local copy of that exchange does not disappear at the end of the conversation, it stays readable in plain text for a period fixed by default. A secret pasted just once into a conversation therefore exists in at least two places afterwards, on the network at the moment it was sent and on the local disk for weeks, which ties back to the rule of never letting a sensitive credential pass through a conversation, developed in the lesson on secrets that must never pass through.

Archiving rather than deleting

A permanent file deletion removes the information with no way back. Moving it to an archive folder keeps the same information available while taking it out of the current working path, and stays reversible at any time. The difference comes down to a single command.

mkdir brouillon_a_ranger
echo note périmée > brouillon_a_ranger\vieux.txt
mkdir _ARCHIVES
move brouillon_a_ranger\vieux.txt _ARCHIVES\vieux_2026-09-02.txt
dir _ARCHIVES

The _ARCHIVES folder now holds the file, still readable, still recoverable, under a dated name that avoids silently overwriting a file with the same name already present in that folder. The same sequence of steps with a deletion instead of a move would make this file permanently lost from the very first command.

An internal document is filtered before it goes out

An internal document is written without filtering precisely because it serves an internal purpose: it names purchase amounts, margins, server paths, people. That same richness becomes a risk the moment the document goes to a third-party formatting tool or to someone outside the organisation. Two separate passes correct this risk. The first is a raw inventory: note every piece of sensitive data present in the document, without yet judging whether it should stay. The second is an adversarial re-read: ask, line by line, what a hostile reader would take from it, a commercial leak, a piece of data that incriminates its author, a shaky claim they could turn against the organisation. A single read that only looks for formatting mistakes does not replace this second pass.

The three reflexes share the same logic: know what leaves, keep what can still be useful, and filter what stays internal for good reason before sending it out.