All the major browser agents already mask secrets in screenshots and use placeholders + redact the dom summaries, I guess the extra scrubbing of faces and other PII could be useful though.

Why not build it as a more general redaction library that can be plugged into any agent loop instead of as a whole agent of its own?