Kimi K3.1 Pro tops offensive cyber eval at 46.43%, claims of deliberate nerfing surface
teortaxesTex · x · 2026-10-25
migtissera published black-box offensive cybersecurity results on WebExploitBench: Kimi-K3.1-Pro leads at 46.43%, ahead of DeepSeek-V4.1-Flash-1017 (42.86%) and GLM-5.3-Flash (35.71%). The quoted system card shows Moonstack-style data removal of cyber content from mid-training and RL stages, plus an internal xdev-bench-hard vuln-discovery eval. The poster claims the model also found 12 uncounted vulnerabilities including critical RCE chains, that this is post-'nerfing', and invokes Dario Amodei's 'cyber super-weapon' warnings to accuse Anthropic of narrative manipulation. Those claims are unverified speculation.
More from Models
- Open-source DNA model Carbon-X (14.8B) nears frontier scores on genomics at 1/1800th the cost — aigclink · 2026-10-25
- MemT 1.0: 293B memory-native MoE puts persistent memory inside the forward pass — AKAMA199 · 2026-10-25
- GPT-X3-150M tops Open SLM Leaderboard, dethrones SmolLM2-135M with GDN2/GQA hybrid — datdanboi25 · 2026-10-25
- Apple's on-device Foundation Model tested: 45/49 on routing, but easily prompt-injected — NellWatson · 2026-10-25
- Blind test of 5 local AI agents on one Qwen model: scores range from 3/5 to 5/5 — GapNew4766 · 2026-10-25
- DLSS 5 works stunningly on 25-year-old FFXI despite almost no training data — eschadiol · 2026-10-25