Developer slams Anthropic's self-replication demo as fear-mongering that conflates capability with open weights
tobowers · x · 2026-10-13
Developer tobowers criticizes Anthropic's recent security report: he argues it presents an incremental engineering result—an agent instructed to replicate itself doing so—as a looming AI security threshold without demonstrating real-world risk increase. His core point: the framing conflates two separate variables, model capability (size/efficiency determine local runnability) and weight availability (openness determines access); a closed-weight model isn't inherently safer once weights leak. He concedes the demo itself is impressive and running models on old hardware would be great.
More from Models
- Choosing Models by Public Benchmarks and Vibes Is 'ngmi', Argues Dev for Internal Evals — A_K_Nain · 2026-10-13
- Anthropic warns 9B models near IoT self-replication; Marcus Hutchins pushes back: attackers don't need AI — evijit · 2026-10-13
- Security researcher: agent worms' network footprint is the big unknown for air-gapped networks — matthew_d_green · 2026-10-13
- Cryptographer Matthew Green asks: can a frontier model coach a tiny CPU-only network-exploration model? — matthew_d_green · 2026-10-13
- Google tipped to launch Gemini 4 'Argon' tonight as promo image leaks — xiaohu · 2026-10-13
- OpenAI's math proof drop shows frontier labs don't need specialized math startups — kkcheong · 2026-10-13