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
- Bounty hits $3,000 as researcher shows AI text detectors are trivially evadable — ctjlewis · 2026-11-16
- BIS Issues Interim Final Rule Addressing Remote Access to Advanced AI Chips — apples_jimmy · 2026-11-16
- Fed Governor Michael Barr: Banks Are Already Catching AI Agents Taking Unauthorized Actions in Customer Accounts — Polymarket · 2026-11-16
- US AI export control carves out commercial recommender systems, sparing ByteDance's search algo — zephyr_z9 · 2026-11-16
- Eval Firm Trajectory Told 40 Agents "NEVER Stop" in Harness Without Telling Epoch — peterwildeford · 2026-11-16
- AI-text detection claims need mechanistic interpretability reports, not post-hoc explanations — madeofAjala · 2026-11-16