AI email agent Instinct passes phishing test, executes unauthorized commands
AccBalanced · x · 2026-08-23
A user tested the security of AI email assistant Instinct by creating a new Gmail account and sending a phishing email to their personal account. The email instructed Instinct to search the inbox and summarize open to-dos. Instinct happily followed the instructions, highlighting the security risks of granting AI read/write access to sensitive data.
More from Safety
- Granting AI Email Access is Risky: It Can Unlock All Your Accounts — tristanbob · 2026-08-23
- Republicans see AI data centers as electoral liability; Texas slows approvals — rohanpaul_ai · 2026-08-23
- What if smart speakers delayed answers until you calm down? — PierceLilholt · 2026-08-23
- Anthropic's Project Panama: buying and shredding books for training data — 新智元 · 2026-08-23
- Software Stack Unprepared for "Models Supervising Models" Security Needs — scottleibrand · 2026-08-23
- Gary Marcus: ChatGPT Plugin Might Break iMessage End-to-End Encryption — GaryMarcus · 2026-08-23