Sentry founder on agent permissions: no secret read access by default

zeeg · x · 2026-11-17

Sentry founder David Cramer weighs in on what permissions AI agents should get: he'd deny any read access to secrets by default, and allow write access only case by case. His mental model: "If Joe accessed my computer, what could he do?" — and an agent adds prompt injection risk on top of Joe.

He argues it depends on the state of the world: exfiltration opportunities are too easy today, so nothing is safe by default. Vendors have a responsibility to protect users who won't make the right decisions but will blame them afterwards — Sentry's approach is to only expose safe credentials like DSNs and SENTRY-prefixed variables.

Original post →

More from coding & agent

coding & agent channel →