Gary Marcus flags prompt injection: AI agents are easy prey for phishing emails
GaryMarcus · x · 2026-10-18
Gary Marcus amplified a Malwarebytes warning that AI agents make users easy targets for phishing: attackers can hide instructions in emails that humans can't read but AI systems process, via prompt injection attacks.
- ShipItClean developer Apollo Raines says the core flaw is that models can't distinguish text to read from instructions to follow.
- The fix isn't user vigilance—vendors should strip hidden content before it reaches the model.
- Users wary of sketchy emails should learn to "distrust their AI assistant" too.
Related event: Security Experts Warn AI Agents Vulnerable to Hidden-Instruction Phishing(2 posts)→
More from AGI Musings
- Teen solves years-old math problem in a week with AI and Lean, writes 15-page paper — IgorCarron · 2026-10-18
- GTA V on an iPhone and weekend vibe-coded games: the new piracy era — aigleeson · 2026-10-18
- South Park underpants-gnome meme skewers the missing step of AGI utopia — Puzzleheaded-King584 · 2026-10-18
- Reddit proposal: offline AI agent 'Payments Mesh' using Bluetooth, signed payments and relayed settlement — Unlucky-Drag8187 · 2026-10-18
- Agentic obliteration is coming: law and medicine first to fall, argues VC — _AustinCalvert_ · 2026-10-18
- MPI professor: I can advise 6 PhDs but only 2-4 AI agent projects at once — Michael_J_Black · 2026-10-18