Default EventScrubber exact-match denylist misses compound secret aliases

[geoffg-sentry]

Defence-in-depth follow-up from VULN-2964. Not a bounty-eligible vuln: default Relay datascrubbing still filters these keys before storage. The SDK-side gap is still worth closing because it is the only layer that prevents transmission.

Problem

EventScrubber.scrub_dict uses exact set membership (k.lower() in self.denylist). Locals capture is on by default. Compound names such as api_token, access_token, auth_token, refresh_token, client_secret, and secret_key therefore leave the process unscrubbed, while password, api_key, private_key, and token do not.

DEFAULT_DENYLIST is commented as stolen from Relay, but Relay's @password:filter matches keys by substring (token, secret, …). The SDK copied the word list, not the matching semantics.

Relevant code:

• sentry_sdk/scrubber.py (DEFAULT_DENYLIST, scrub_dict)
• sentry_sdk/client.py (default scrubber install, scrub_event)
• sentry_sdk/consts.py (include_local_variables defaults True)

Suggested fix

Either is fine:

1. Match keys by substring / root terms (token, secret, auth, password, key, credentials), similar to _SENSITIVE_DENYLIST in sentry_sdk/data_collection.py, or
2. Keep exact match and add the missing literals.

Substring matching over-scrubs (tokenizer, keyboard, author). Relay already uses anchors (^otp$, ^two[-_]factor$) for that reason. Expanding the literal list is equally defensible. Silently diverging from the list we copied is not.

Related, not this ticket

Enabling _experiments={"data_collection": ...} uninstalls EventScrubber (client.py ~355–368) and never key-filters frame["vars"]. That path is worse than default, including for password / api_key. Raise it before data_collection becomes default.

Related: PY-1691 / sentry-python#4441.

[linear-code]

PY-2791

[ericapisani]

Regarding this comment here:

Enabling _experiments={"data_collection": ...} uninstalls EventScrubber (client.py ~355–368) and never key-filters frame["vars"]. That path is worse than default, including for password / api_key. Raise it before data_collection becomes default.

This is correct in that the EventScrubber is uninstalled and therefore doesn't key-filter, but this isn't entirely correct that there's no scrubbing on this value whatsoever.

Sensitive data filtering on the frame now takes place in the serialize_frame method (see here).

That being said - the default setting for the data_collection property controlling this scurbbing is true as defined in the spec. As a default, this means that no scrubbing is applied, as the sensitive deny-list was only defined to apply to key-value data.

I'm going to raise this with the spec designers as we support key-value collection behaviour here, but it's currently a configurable option rather than the default (think maybe this should be the other way around, where a denylist is the default and people can opt to collect everything using True as the config value).