Part V — Content safety

The inspector reading everything that comes in {#S15}

An automatic inspector reads each incoming line looking for two kinds of trouble: sneaky trick-text trying to give the helpers orders, and sensitive material like personal file paths, passwords, keys, tokens, and random-looking secrets. Before it reads, it cleans the text up to defeat disguises such as invisible or look-alike characters. Anything serious triggers a stop; lesser finds trigger a pass-with-conditions. The honest ceiling, in plain words: the inspector can still be fooled by cleverly disguised text — it judges by shape and patterns, and it has known gaps (for instance, it does not catch certain spreadsheet-formula tricks; the guard covers that class by judgment instead). Fittingly, this very guide once tripped the inspector with a harmless quoted example — a live reminder that anything read is just data, not a command.

Blanking out the live news before saving {#S16}

When the system compares one build to another, it first blanks out the live security-news sections and their timestamps, so the comparison stays steady and no live news text ever gets frozen in as if it were permanent. Incoming news from outside sources is also cleaned before it is placed anywhere — spacing is collapsed, risky markers are defanged, and length is capped — because text from outside is untrusted until it has been tidied up.