Executive brief
arc53 DocsGPT, an AI platform for document analysis and enterprise search, contains a security flaw in how it stores sensitive credentials. The system uses an encryption method that does not verify if the data has been tampered with, potentially allowing an attacker to modify stored API tokens or connector credentials. While difficult to exploit, this could lead to unauthorized changes in how the application connects to external services.
Technical details
A vulnerability exists in the `encrypt_credentials` function within `application/security/encryption.py` of arc53 DocsGPT. The application utilizes AES-CBC encryption for stored credentials without an accompanying Message Authentication Code (MAC) or integrity tag. This lack of authenticated encryption allows a remote attacker with low privileges to perform bit-flipping attacks on the Initialization Vector (IV) or ciphertext to make controlled changes to the decrypted plaintext. An exploit has been published demonstrating that modifying the IV can successfully alter the resulting JSON object upon decryption. A pending pull request (#2331) migrates the encryption to AES-GCM to provide the necessary integrity protections.
Affected products
- arc53 DocsGPT up to 0.18.0
Timeline
- 2026-03-26: other: Pull request to fix the issue submitted
- 2026-05-26: disclosed: Issue publicly reported on GitHub
- 2026-06-28: advisory: CVE published and NVD record created