Palisade Research shows that AI agents can hack remote computers, copy themselves onto them, and form replication chains. In one year, the success rate jumped from 6 to 81 percent. The researchers expect remaining barriers to fall as models get better at hacking.

METR can barely measure Claude Mythos Preview with its current test suite. Only five out of 228 tasks cover the relevant capability range. Meanwhile, Palo Alto Networks reports that frontier models autonomously chain vulnerabilities, shrinking the time from initial access to data exfiltration to just 25 minutes. Evaluation methods are growing more slowly than the models themselves, and that may be the bigger problem.

ByteDance is raising its planned AI spending for 2026 to over 200 billion yuan (roughly $30 billion), at least a 25 percent jump from earlier plans. The TikTok parent is increasingly turning to Chinese chips. Still, the figure looks modest next to the $725 billion that Google, Amazon, Microsoft, and Meta are planning to spend combined.

OpenAI doubled GPT-5.5's list price compared to GPT-5.4, claiming shorter responses would offset the increase. An OpenRouter analysis of real usage data tells a different story: actual costs rose 49 to 92 percent depending on input length. Anthropic hiked Opus 4.7 prices too—and with both companies eyeing IPOs, the trend is unlikely to stop.

A study by researchers from the MATS program, Redwood Research, the University of Oxford, and Anthropic examines a safety problem that grows more pressing as AI systems become more capable: "sandbagging," where a model deliberately hides its true abilities and delivers work that looks adequate but is intentionally subpar.