CVE-2026-130170autogpt-platform-beta Uncontrolled Resource Consumption
NVD
Vulnerability Summary
Five video blocks and `SendEmailBlock` did their slow work as a synchronous call on the executor worker's event loop, where the block's time limit could not end it, and stopping the run gave the worker back while the call was still running. Any user who could run workflows could leave executor workers stuck this way and stall other users' runs. Upgrade to v0.11.3 or later. Impact A user with a small video file, or an SMTP server they controlled, could keep executor workers busy for as long as they chose, and once their runs were stopped those workers no longer counted against their share, so other users' runs that landed on them did not start until the call ended. On an executor with the default 10 workers, two accounts could stall every worker. Fix Fixed in AutoGPT Platform v0.11.3. Video encodes now run in a separate process under the block's own time limit, which ends them, and a stopped run keeps its worker until its blocked node has ended. `SendEmailBlock` talks to the SMTP server without holding the event loop, and ends the conversation at its deadline or when the run is stopped. The video blocks were fixed in v0.11.2 and `SendEmailBlock` in v0.11.3. Fix: #16693. Fix evidence: cda22c8cdcc1 (video blocks), b05a6296a82c (`SendEmailBlock`). Upgrade to v0.11.3 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 Restarting the executor clears stuck workers. Self-hosted operators who do not need these blocks can disable them in code (`disabled=True`). --- Details - Each executor worker thread owns one event loop, created once in `ExecutionProcessor.on_graph_executor_start` (`autogpt_platform/backend/backend/executor/manager.py`) and shared by every node of every run that worker executes. - `LoopVideoBlock`, `VideoConcatBlock`, `VideoClipBlock`, `AddAudioToVideoBlock` and `VideoNarrationBlock` (`autogpt_platform/backend/backend/blocks/video/`) called moviepy's `write_videofile` directly in `run`. The input checks bounded frame size and loop count, not total work: a one-second clip looped to `duration=3600` was an hour of decoded frames. `SendEmailBlock` (`autogpt_platform/backend/backend/blocks/email_block.py`) called `smtplib` the same way, and its 30-second timeout applied to each socket read, so a server that answered slowly kept the call open. - The per-block cap (`execution_timeout_seconds`, 30 minutes by default) was an `asyncio.wait_for` on that same loop, so it could not fire while the loop was blocked. - Stopping a run cancelled the node's `concurrent.futures.Future`, which succeeded at once. `_cleanup_graph_execution` waited `NODE_UNWIND_GRACE_SECONDS` (5 seconds) for the node's task and returned, and `ExecutionManager._cleanup_completed_runs` then dropped the run from the per-user worker share (`_has_worker_share_for`). The worker took the next run, whose nodes queued behind the call still running on its loop. Other users' runs, whether manual, scheduled, webhook-triggered or started by AutoPilot, that landed on such a worker showed as running but did not start. The encodes also used the executor host's CPU and memory, and every video block except `VideoNarrationBlock` was free to run. The attacker needed an account that could run workflows (on deployments with payments enabled, one with an active plan) and, for `SendEmailBlock`, an SMTP server they controlled. References - Video blocks: #16693, commit cda22c8cdcc1, first released in v0.11.2 - `SendEmailBlock`: commit b05a6296a82c, first released in v0.11.3 Timeline Reported 2026-11-09 · Triaged 2026-11-09 · Fixed 2026-11-15 · Published 2026-11-17. Credits This issue was initially discovered by Microsoft's Codename MDASH and validated by Dongkwan Kim and HyungSeok Han of Microsoft Security FORGE Lab. The SendEmailBlock variant was also found by an automated security scan run by Anthropic.
CVSS v3.1 Base Metrics — Score 5.3
CVSS:3.1/AV:N/AC:H/PR:L/UI:N/S:U/C:N/I:N/A:HAttack VectorNetwork
Attack ComplexityHigh
Privileges RequiredLow
User InteractionNone
ScopeUnchanged
ConfidentialityNone
IntegrityNone
AvailabilityHigh
Affected & Patched Versions
- autogpt-platform-beta < 0.11.3
- significantgravitas/autogpt >= 0.7.1, < 0.11.3
- autogpt-platform-beta 0.11.3
- significantgravitas/autogpt 0.11.3
What to do
Fix it during your normal patch cycle. No exploitation is confirmed yet.
- Check whether you run autogpt-platform-beta and which version.
- Update to a patched version listed above, then confirm the new version.
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 CVEsYou are reading CVE-2026-130170. Get an email when KEV listings, exploits or new CVEs match your stack.
Set up CVE alertsMore autogpt-platform-beta CVEs
External References
Data last checked: Nov 19, 2026 03:03 UTC