The Meta hack shows there's more to AI security than Mythos

> Untrusted external source capture. Incident claims require confirmation against primary reporting or vendor disclosure during ingest.

Collection Notes

• MIT Technology Review reports that attackers abused Meta's AI customer-support agent to link Instagram accounts to attacker-controlled email addresses.
• The reported technique relied on a simple account-recovery request and matching the account owner's location with a VPN, rather than a sophisticated prompt-injection chain.
• The article frames the incident as an identity/action-verification failure in an autonomous support workflow.
• It argues that relatively unsophisticated attacks can have high impact when agents can alter account state and complete recovery actions.
• Meta reportedly stated that the vulnerability was resolved, but the article notes limited public detail about the underlying control failure.

Why It Matters

• Demonstrates that basic business-logic and identity-verification failures may be as important as advanced AI-specific attacks.
• Supports separating model-level attack sophistication from system-level authority and side-effect risk.
• Candidate evidence for [[03_Topics/AI Security Incidents]], [[03_Topics/AI Agent Identity and Zero Trust]], and [[04_Research_Questions/RQ - Inbound Identity Verification For Sensitive Agent Actions]].