Executive brief
AutoGPT, a platform for managing AI agents, is vulnerable to a denial-of-service attack that can crash the system. By repeatedly requesting screenshots of numerous web pages, a malicious user can fill up the server's disk space because the system does not limit how many images are stored or how many times the process can loop. This can lead to a total service outage, preventing the platform from functioning for all users.
Technical details
A denial-of-service vulnerability exists in AutoGPT due to uncontrolled resource consumption (CWE-400). The `ScreenshotWebPageBlock` stores captured screenshots in a temporary directory without enforcing disk space quotas, while the `StepThroughItemsBlock` allows for infinite iterations of these capture tasks. An attacker can exploit this by providing a large list of URLs to screenshot, eventually exhausting the host's disk space. Furthermore, because agents can be configured with long-running timers, the temporary files may not be cleared for an extended period, leading to a persistent DoS state. The issue is resolved in version 0.6.63.
Affected products
- Significant-Gravitas AutoGPT < 0.6.63
Timeline
- 2026-06-11: advisory: GitHub Security Advisory published
- 2026-06-18: disclosed: CVE published to NVD