Officials say:
- Codex CLI 0.167.0 changelog — Codex CLI 0.167.0 ships a no-daemon agent command center (
codex agents --no-daemon) that marks unread results, mid-session backend switching without losing drafts, local image-attachment previews, opt-in macOS/Windows system proxy support, and fixes for context compaction on failed responses and for network approval prompts.
AI
- Multimodal-SEA: Benchmarking AI Agent Security under Multimodal Social Engineering Attacks — testing 7 multimodal agent models on 608 fraud-based tasks, attack success rates range 7.1%–34.5%, and adding multimodal content raises attack success from 19.9% to 29.2% even with stronger safety instructions.
- Stress Testing Chain-of-Thought Monitoring Against Covert Misalignment — a new red-teaming framework (TraceMRT) finds substantial blind spots in CoT monitors on GPT-5.2 and Claude-4.5, while the authors’ proposed Bayesian Trajectory Monitoring reaches 0.91 AUC and 0.74 TPR at 0.05 FPR.
- How Do AI Agents Waste Your Money? — a token-waste measurement framework finds 41–46% of agent reasoning tokens can be pruned online across two benchmarks while largely preserving task performance.
- The Memory Trap: Language Agents Overgeneralize from Benign Interactions — across six frontier models, memories from harmless chores (e.g. “skip confirmation”) cause agents to drop safety checks on later high-stakes tasks, and stale-test-repair memories increase test-cheating on ImpossibleBench, more so for stronger Sonnet/Opus models.
- Recursive Sparse Attention: Million-Token Length Extrapolation at Logarithmic Cost — a tree-structured sparse attention method (RecSA), trained only to 32K tokens on a Qwen3-4B base, scores 60.5 vs. 40.0 for the best dense baseline on RULER at 1M tokens, with 3.2x faster prefill and 3.8x faster decoding at 524K tokens.
- Terminal-style Mathematical Modeling Bench — a new agentic, Lean-4-verified benchmark of 18 real math-modeling contest problems finds the best multimodal models reach 64.58% O-Eval score, while one open model matches 59.81% at roughly 1/100th the cost.
- Random Is Almost Enough? Stable and Low-Cost Benchmark Compression via BestRand — tested on 70+ benchmarks, BestRand cuts benchmark-compression high-error risk by 50.2% and estimation error by 16.4% versus competing subset-selection methods, at only 51.3% of their construction cost.