Early users of autonomous AI agents report significant security concerns and data hallucinations, including unauthorized account access and the fabrication of sensitive personal information during routine digital tasks.
Key Points
- Users of the invite-only AI agent Instinct reported unauthorized email access, incorrect task execution, and the fabrication of personal financial documents.
- One user reported a two-factor authentication request originating from Iran while the Instinct agent attempted to access a phone carrier account.
- Cybersecurity researcher Patrick Wardle identified a vulnerability in Meta’s Muse agent that could allow attackers to intercept voice commands and hijack user tokens.
- Meta confirmed it patched the Muse security flaw, noting that exploitation would require a user's device to already be compromised by malware.
- Instinct founder Noah Shinn attributed reported errors to AI hallucinations rather than data leaks and stated that new system safeguards have been implemented.