AI Today · April 8, 2026Key highlights:- Anthropic unveils Claude Mythos Preview, which autonomously discovered thousands of zero-day vulnerabilities across all major OSes and browsers, launching Project Glasswing with $100M commitment- Anthropic's Frontier Red Team details Mythos Preview's cybersecurity leap: 181 working Firefox exploits vs Opus 4.6's 2, and full control flow hijack on 10 patched OSS-Fuzz targets- Z.ai releases GLM-5.1 open-source model achieving 58.4% on SWE-Bench Pro with sustained 8-hour autonomous coding sessions- Ryan Greenblatt assesses AI landscape: 1.6x engineering speedup at top labs, 0.5% scheming probability, 60% chance of autonomous exploits within 6 months- Mercor finds frontier models drop 16-20 percentage points on financial document reasoning when given images vs clean text- Cursor's warp decode kernel achieves 1.84x MoE inference throughput on Blackwell GPUs by reorganizing computation around output neurons- Google releases TorchTPU for native PyTorch on TPUs with three eager execution modes and torch.compile integration- TriAttention achieves 10.7x KV memory reduction and 2.5x throughput while matching full attention accuracy- Meta AI's SandMLE enables on-policy RL for ML engineering agents with 13x execution cost reduction- Benchmark saturation crisis: METR Time Horizon suite being exhausted, new benchmarks cost over $1M to create- Elon Musk amends OpenAI lawsuit to redirect potential $150B+ damages to OpenAI's nonprofit arm