Anthropic's next AI model Mythos can hack you...

Summary

This video catches you up on the biggest AI news from the past week, starting with the accidental leak of Anthropic's upcoming model, Claude Mythos (also referred to as Capybara or MEOS). Anthropic briefly published a pre-release article about Mythos before quickly deleting it, but not before it was saved and covered by Fortune magazine. What stands out about Mythos is its positioning above Claude Opus — this is not an incremental update but a significant leap in capability. One of the most notable signals of its power is that Anthropic is proactively warning cybersecurity professionals to patch their systems before the model releases, implying it will be capable of identifying and exploiting vulnerabilities in ways current models cannot. The Cursor/Claude team is also suspected to be using Mythos internally, which may explain their rapid feature shipping pace — even Dario Amodei has publicly noted that his team no longer writes code manually. On a less positive note, Anthropic has quietly reduced usage limits for Claude subscribers without clear communication, frustrating researchers and power users. There is no transparent usage dashboard like competitors such as MiniMax offer, leaving users guessing why their workflows are suddenly failing. This is speculated to be tied to Anthropic redirecting compute toward training Mythos. Shifting to the Chinese AI space, Zhipu AI released GLM 5.1, which shows strong improvements in coding and agentic tasks. You can use it through Claude Code, and even GLM's own team openly promotes Claude Code over building their own coding tool. Early benchmarks on WildBench place it near Xiaomi's Mimo, though testing has been hampered by severe server congestion since launch. Finally, the video teases coverage of a rising open-source agent called Hermes, which has been gaining significant traction online.