Skip to content

Agent pool Postgres DSN is logged in cleartext, including the password, while the auth DSN is redacted #11991

Description

@66656c6c6f7574

LocalAI version

v4.9.0 (localai/localai:latest-gpu-nvidia-cuda-12, revision f7ad3f70eb5d8a0ddf80e08557f0d7df28cf032e)

Describe the bug

When the agent pool uses the Postgres vector engine, LocalAI logs the full
LOCALAI_AGENT_POOL_DATABASE_URL at INFO level, including the database
password in cleartext, every time it opens a collection.

The auth database URL is redacted correctly in the same log stream, so the
redaction helper already exists. It just is not applied on this path.

Redacted (correct), from LOCALAI_AUTH_DATABASE_URL:

INFO Auth enabled database="postgresql://%2A%2A%2A:%2A%2A%2A@postgres:5432/localai_auth?sslmode=disable"

Not redacted, from LOCALAI_AGENT_POOL_DATABASE_URL (password replaced by hand
before pasting here):

INFO PostgreSQL collection collectionName="AML" databaseURL="postgresql://localai_user:REDACTED-PASSWORD@postgres:5432/localai?sslmode=disable"

One line is emitted per collection on every start, so the credential is
repeated and persists in whatever collects container logs. In a Docker
deployment these sit on disk under the json-file driver and are readable by
anyone who can read container logs, for example through a management UI,
without any access to the environment or the compose file.

To Reproduce

  1. Set LOCALAI_AGENT_POOL_VECTOR_ENGINE=postgres and
    LOCALAI_AGENT_POOL_DATABASE_URL=postgresql://user:password@host:5432/db?sslmode=disable
  2. Have at least one agent collection defined.
  3. Start LocalAI and read the log.

The PostgreSQL collection line contains the password verbatim.

Expected behavior

The agent pool DSN should be redacted in logs the same way the auth DSN
already is, or the log line should print only the host, port and database
name rather than the whole URL.

Additional context

Not configurable from the outside: the value has to be a DSN, so there is no
way for an operator to avoid this short of not using the Postgres vector
engine. Rotating the password does not help, because the new one is logged
too.

Suggested fix: apply the same redaction used for the auth DSN to the agent
pool DSN before logging.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions