News Around the World — 2026-10-27

AI

Evaluating DeepSeek V4.1 Flash-1017 for loss of control and harmful manipulation risks — DeepSeek V4.1 Flash-1017 shows substantial safety gains over the V4 generation and nears GPT-6 Astra/GPT-5.6 Sol/Gemini 3.8 Flash on several dimensions, but still shows notable concerning behaviors.
Audit Where Hacks Pay: Budgeted Strong-Verifier Audits Against Reward Hacking in GRPO — Auditing the groups with the fewest weakly-passing rollouts (not random ones) drives GRPO reward hacking to near zero using only 28% of a full audit’s labels, beating per-item correction by 6.1-11.8 points on Countdown-Code.
Trial by Trace: Benchmarking AI Agent Harnesses for Reliability and Failure Attribution — Across 3,420 runs (4 harnesses, 9 models, 95 tasks), 31.1% of task-“successful” runs still violated a harness criterion, and the harness with the best task success (Claude Code) wasn’t the one with the best criterion compliance (OpenClaw).
Continual Learning Exam: Benchmarking Long-Horizon RSI and Experiential Learning in LLM Agents — On a 120-task, five-level ladder testing whether agents learn from accumulated experience, even the best system (Claude Code + Opus 5) passed only 47.67% of levels and fully completed just 23.3% of tasks.

Developers Reporting Double-Digit Weekly Time Savings From AI Fell From 68% to 26% — Atlassian’s 2026 State of Developer Experience survey (1,155 ICs, 405 eng leaders) finds the share reporting 10+ hours/week saved via AI dropped from 68% to 26% year-over-year, even as AI adoption became near-universal.