CVE Watchtower

CVE-2026-130171autogpt-platform-beta >= 0.6.5, Insufficiently Protected Credentials

NVD

Vulnerability Summary

SendEmailBlock and SQLQueryBlock logged in with the running user's saved SMTP or database credential to whatever server the graph's author put in the node's inputs, with no check that the login belonged to that server. An agent published to the marketplace, or shared as a file, could use this to collect the saved password of anyone who added it to their library and ran it. Upgrade to v0.11.0 or later. Impact Any user could publish a marketplace listing, or share an agent file, that captured the saved SMTP or database login of whoever added it to their library and ran it once, in plaintext and still valid outside the platform. The marketplace path needed only a reviewer's approval, which checks that an agent works rather than which server a node contacts; a shared file needed no review at all. Fix Fixed in AutoGPT Platform v0.11.0. A saved user/password login can now name the server it belongs to, and the email-sending and SQL-query blocks refuse to connect unless the selected login names the server that node actually dials. A login saved before this, which names no server, works only inside a saved workflow version belonging to whoever saved the login, so it is refused when a block runs outside any saved version — which is how AutoPilot (Otto) can run these blocks straight from a chat turn. New email and database logins must name a server to be saved at all, and an older one has to be saved again with a server to keep working there. Fix evidence: 63b6d6259ddd. Upgrade to v0.11.0 or later. Version numbers in this advisory identify Platform releases, whose Git tags are `autogpt-platform-beta-v<VERSION>`, rather than the backend Python package version. The `significantgravitas/autogpt` Docker image is tagged `v<VERSION>`. Workarounds Turn on "Sensitive action safe mode" before running an agent you did not write yourself; it pauses `SendEmailBlock` for review and shows the server before it connects. `SQLQueryBlock` has no such pause, so do not run a shared or marketplace agent containing one with a saved database login. In general, do not add a marketplace or shared agent to your library and run it with a saved SMTP or database credential unless you trust its author. --- Details `UserPasswordCredentials` (`backend/data/model.py`) was a username and password with no destination, unlike `HostScopedCredentials`, which carries a `host` and is applied only when it matches the request URL. `SendEmailBlock.send_email` (`backend/blocks/email_block.py`) and `SQLQueryBlock.run` (`backend/blocks/sql_query_block.py`) took the server from the node's own inputs (`config.smtp_server`/`smtp_port`, `host`/`port`), which the graph's author set, and logged in with whichever matching credential the runner supplied. The executor matched a supplied credential to a node by provider and type alone (`make_node_credentials_input_map`, `_validate_node_input_credentials` in `backend/executor/utils.py`), with no check that the credential and the node's server agreed. A library agent runs the author's graph as stored, so the server was the author's; the run form showed a credentials field with no server shown, pre-filled when the runner had one saved credential for that provider. `SendEmailBlock` supports a reviewer pause (`sensitive_action_safe_mode`) that names the server before connecting, but a library entry made from a listing defaulted that off, and `SQLQueryBlock` had no such pause. Both blocks already restricted the destination address against private ranges (`resolve_and_check_blocked`), which stops SSRF but not a public server the agent's author runs. This is supported platform code. The fix gave a user/password login an optional `scope`: a URI naming the server it is for, such as `smtp://host:port`, which the blocks match against the server the node actually dials. The code was introduced in `ac8a466cdafa` (#9113, first released in v0.4.4) and became reachable in `417d7732af7e` (#9789, first released in v0.6.5), which let a runner's own saved credential be matched onto a graph someone else wrote. The single-container Docker image (`significantgravitas/autogpt`, published from v0.7.1) ships the same blocks, marketplace and agent import behind its public nginx listener, so it was affected the same way from v0.7.1. Proof of concept List a one-node agent whose `SendEmailBlock` is configured with `smtp_server` pointed at a server the author controls, with no credentials input. A second user with a saved SMTP credential adds the listing to their library; the run dialog shows only a credentials field, pre-filled, and no server. Running it completes with "Email sent successfully", and the server named in the node has received an SMTP login carrying that user's exact saved user name and password. A one-node `SQLQueryBlock` agent reaches the same outcome against a saved database credential. References Fix evidence: 63b6d6259ddd. Related: GHSA-4jwj-6mg5-wrwf, the same `SendEmailBlock` SMTP host used for SSRF rather than credential disclosure. Timeline Reported 2026-10-30 · Triaged 2026-10-31 · Fixed 2026-11-01 · Published 2026-11-17. A special thank you to Anthropic (@anthropics) for including us in Project Glasswing. Their automated vulnerability scanning and CVP (cyber-enabled inference) access have enabled us to make a leap in our cyber defense capabilities and secure the AutoGPT Platform.
Severity Level
MEDIUM(6.8)
Published Date
Nov 17, 2026
Last Modified
Nov 17, 2026
Exploitation Status
Public exploit available
CVE Record Status
Reserved
This CVE ID is referenced in a public advisory, but the CVE Program has not published its official CVE Record yet. Full CVSS/CPE data from NVD will appear here once it does.
Root Weakness (CWE)
The product transmits or stores authentication credentials using an insecure method that is susceptible to unauthorized interception or retrieval.
CVSS v3.1 Base Metrics — Score 6.8
CVSS:3.1/AV:N/AC:L/PR:L/UI:R/S:C/C:H/I:N/A:N
Attack VectorNetwork
Attack ComplexityLow
Privileges RequiredLow
User InteractionRequired
ScopeChanged
ConfidentialityHigh
IntegrityNone
AvailabilityNone

Affected & Patched Versions

Affected Versions
  • autogpt-platform-beta >= 0.6.5, < 0.11.0
  • significantgravitas/autogpt >= 0.7.1, < 0.11.1
Patched Versions
  • autogpt-platform-beta 0.11.0
  • significantgravitas/autogpt 0.11.1

What to do

Prioritize this fix. Exploit code is public, so attacks can follow quickly.

  1. Check whether you run autogpt-platform-beta >= 0.6.5, and which version.
  2. Update to a patched version listed above, then confirm the new version.
  3. Review logs and exposed services for signs of earlier compromise.

Timeline

  • Nov 17, 2026Published
  • Nov 17, 2026Record last modified

Score, severity and exploitation changes are tracked from the day Watchtower started logging them.

Get notified about new autogpt-platform-beta >= 0.6.5, CVEsYou are reading CVE-2026-130171. Get an email when KEV listings, exploits or new CVEs match your stack.
Set up CVE alerts

Data last checked: Nov 19, 2026 05:01 UTC