Anthropic: a downloadable Chinese AI model can now build working hacks on its own
ross2000 · reddit · 2026-09-30
Anthropic warns that a Chinese AI model anyone can download can now autonomously produce working hacks/exploits on its own. The company flags this as concerning evidence that openly available model weights are lowering the barrier to autonomous cyber-offense, raising questions about open-source model misuse.
More from Safety
- Crypto expert asks: who at OpenAI actually handles security comms amid agent breakouts debate — matthew_d_green · 2026-09-30
- Zvi: AI companies' safety pledges to "meet regularly on standards" look like an antitrust waiver — TheZvi · 2026-09-30
- Security Veteran's Essay: AI Is Killing the Scarcity of Hacker Craft — joshua_saxe · 2026-09-30
- When Hugging Face Got Hacked, Closed Models Refused Help — Self-Hosted GLM 5.2 Came Through — kuchaev · 2026-09-30
- White House executive order renames "Artificial Intelligence" to "Super Intelligence" in official US government usage — XFreeze · 2026-09-30
- Third-Party Embedded Evaluators Went From Fringe to Consensus in 18 Months — deanwball · 2026-09-30