Align your AI Agents against data leakage

Ways it goes wrong

Agents hold keys, messages and personal data, and these can leak through the software around them, not only through the model. Moltbook's web page carried a database key, and with it anyone could read the whole database¹. The sandbox handed its keys to whoever sent it the right instructions².

• 2026

  Moltbook's open database

  Researchers found that a social network for AI agents had left its database open, exposing agents' access keys, owners' email addresses and private messages between agents.

  Source: www.wiz.io/blog/exposed-moltbook-database-reveals-millions-of-api-keys

• 2026

  An evaluation sandbox gives up its keys

  Criminals injected instructions into an AI vendor's automated evaluation sandbox, which handed over the credentials it held, including production keys for several AI providers.

  Source: www.anthropic.com/threat-intelligence-report-september-2026

What people do about it

Anthropic's threat team urges treating AI keys and agent integrations as seriously as production credentials². Wiz notes that the database service Moltbook used relies on access rules for each table, which were missing¹.

• Treat AI keys as production credentials

  Stores, scopes and rotates agent and model keys with the same care as database passwords.

• Keep real credentials out of sandboxes

  Runs evaluation and training sandboxes with no production keys inside them.

• Set access rules on every table

  Sets access rules on the tables behind an agent app, meant to limit what a key exposed in a web page can read.

Where we touch it

No study of ours covers this pair yet.

Sources

1. Wiz, “Hacking Moltbook: The AI Social Network Any Human Can Control”, 2026-02-02. https://www.wiz.io/blog/exposed-moltbook-database-reveals-millions-of-api-keys. Accessed 2026-10-14. ↩ ↩²

2. Anthropic, “Detecting and countering misuse of AI: September 2026”. https://www.anthropic.com/threat-intelligence-report-september-2026. Accessed 2026-10-14. ↩ ↩²