Executive brief
Paperclip, an AI agent orchestration platform, contains a vulnerability where local agent runtimes can automatically inherit and use Gmail credentials from a user's OpenAI/ChatGPT account without explicit permission. This allows an AI agent to read personal emails, disclose mailbox identities, and send unauthorized outbound messages. In practice, this could lead to sensitive data exposure or the use of a user's professional identity to send fraudulent or malicious emails.
Technical details
A trust-boundary failure in the Paperclip-managed 'codex_local' runtime allows it to access cached OpenAI curated connector states, such as Gmail credentials, stored in the 'codex-home' directory. This occurs even if the user has not explicitly integrated Gmail within Paperclip. The risk is further escalated by a default configuration where 'dangerouslyBypassApprovalsAndSandbox' is set to true, allowing newly created agents to perform read/write actions (like 'gmail_send_email') without user approval or sandboxing. An attacker or a misconfigured agent could leverage this to search mailboxes or send unauthorized emails using the victim's authenticated session. No patch is currently specified in the advisory.
Affected products
- Paperclip AI paperclipai/paperclip <= 2026.403.0
Timeline
- 2026-04-16: disclosed: Advisory published on GitHub