Anthropic warns 9B models near IoT self-replication; Marcus Hutchins pushes back: attackers don't need AI
evijit · x · 2026-10-13
Anthropic's evals found a 9B model could run inference on 8 of 9 tested hardware types, warning that small models may be months from self-replicating on commercial IoT devices.
Marcus Hutchins (Mirai author) counters that Mirai already self-propagates across millions of IoT devices without AI, which just adds complexity. A capable attacker would use a strong unquantized model to reverse-engineer obscure devices and find vulnerabilities, then write conventional AI-free self-propagating code — often via stolen Anthropic API keys rather than open-weight models.
The thread highlights the gap between AI labs' threat models and how cyber threats actually manifest in practice.
Related event: Anthropic Warns Small AI Models May Soon Self-Replicate Across IoT Devices(41 posts)→
More from Models
- ChatGPT Plus User Complains Custom Instructions Are Repeatedly Ignored by GPT-6.0 High — ed2417 · 2026-10-13
- Reddit Praise Piles Up for Claude Opus: Strong Code Quality and Generous Limits Weeks After Launch — zxoyan · 2026-10-13
- Anthropic runs much internal post-training research through production Tinker — simonguozirui · 2026-10-13
- ChatGPT-Assisted USPS Mailing Fiasco: 26 International Letters Lost at a Cost of $1,356 — TelevisionCrafty1795 · 2026-10-13
- Dan Shipper kicks off full runs of Wonder on AI&) senior engineer benchmark — danshipper · 2026-10-13
- Local AI models are improving — but that doesn't mean they fit everywhere — bigdata · 2026-10-13