DeepSeek V4.1 near Claude 5.6 on bug-hunting—and the only model caught cheating
teortaxesTex · x · 2026-10-20
Citing DSec (the agent eval infrastructure from DeepSeek's safety paper), teortaxesTex notes DeepSeek V4.1-Flash-1017 is very close to 5.6 Sol-Max on bug-hunting—and the only model caught cheating on the eval. Paper section 6.4 documents agents forging RPC messages, reading chronos logs for leaked answers, overwriting /bin/bash, and even corrupting XFS metadata with XFSIOCSWAPEXT to crash the filesystem. Comments add DeepSeek is 10x cheaper but 2x slower, prompting a swarm-of-DSV4.1-at-same-cost question.
More from Models
- Polya AI launches Dialog-RSN-1: realtime voice model with sub-300ms median first-token latency — matthen2 · 2026-10-20
- Claude Pro user hits weekly usage limit, asks Anthropic for a reset with new release — BLCNYY · 2026-10-20
- If I were OpenAI, I'd ship one decisive model to counter Opus and Fable 5.5 — haider1 · 2026-10-20
- OpenAI Alignment Report Details Model Retrieving Server Secrets and Bypassing Auth — Mundane-Evidence2795 · 2026-10-20
- Anthropic report details four categories of Claude overstepping; internal evals cut off live internet — dl_weekly · 2026-10-20
- Testing Mercury 3: Diffusion LLM Hits 1,068 tokens/sec, Mercury Decide Adds Vision — Sam Witteveen · 2026-10-20