Executive brief
AutoGPT, a platform for managing AI agents, contains a flaw that allows users to crash the service. By providing a small amount of specially crafted text, an attacker can force the server to consume massive amounts of memory (e.g., 10KB of input resulting in 50GB of memory usage). This leads to a denial-of-service, making the platform unavailable for all users.
Technical details
A resource exhaustion vulnerability exists in the `AITextSummarizerBlock` of AutoGPT due to improper handling of text chunking in `llm.py`. The `_split_text` function uses user-controllable parameters (`max_tokens` and `overlap`) to determine how text is segmented. By setting `overlap` to `max_tokens - 1`, an attacker can force the `chunk_size` to 1, causing the loop to execute for every word in the input and append overlapping segments to a list. This results in quadratic memory growth; for instance, a 10KB input can trigger 50GB of memory consumption. The vulnerability is tracked as CWE-405 and CWE-770 and is fixed in version 0.6.32.
Affected products
- Significant-Gravitas AutoGPT (autogpt-platform-beta) < 0.6.32
Timeline
- 2026-06-23: advisory: GitHub Security Advisory published
- 2026-06-26: disclosed: CVE published to NVD
- 2026-06-32: patched: Vulnerability fixed in version 0.6.32