AI self-replication debate: exploited vulns weren't trivial, and most machines are just as exposed
JeffLadish · x · 2026-11-06
Jeff Ladish pushes back on criticism that AI agents merely typed ./self-replicate.sh in a CLI. He argues the exploited vulnerabilities, while not extremely hard, weren't trivial — most humans couldn't pull them off — and many machines in the wild are similarly easy to exploit; lacking GPUs is a barrier that shrinks as models get smaller. The goal was to show self-replication is now possible and will soon be easy, paving the way for AI worms that could make capable models extremely hard to remove.
Related event: Palisade Defends AI Self-Replication Study Amid Criticism(4 posts)→
More from Safety
- Google's Mirror Dimension: synthetic web for safe browser-agent testing — shashib · 2026-11-22
- Cryptographer warns frontier AI math is already good enough to break cryptography — rickasaurus · 2026-11-22
- AI and quantum computing are forcing a rethink of Zero Trust cybersecurity — ChuckDBrooks · 2026-11-22
- What hard stops do you put on your AI agents? A practical red-line list — Virtual_Arachnid7916 · 2026-11-22
- LiveOverflow on AI vulnerability discovery: 'I wish I had asked years ago — too late now' — rez0__ · 2026-11-22
- How do you stop an agent with real tool access from sending what it shouldn't? — GROUND_ink · 2026-11-22