AI Roundup: Claude Max Plans Become Max and Ultra, Anthropic Automates Misalignment Red-Teaming

Claude Max plans renamed to Max and Ultra

Anthropic renamed its Claude Max subscription tiers in an October 30 post from the official @claudeai account:

• Max 5x is now Max, still $100 a month.
• Max 20x is now Ultra, still $200 a month.

Anthropic says subscribers don’t need to do anything. It’s also removing the five-hour limit for verified Ultra accounts, starting with most existing subscribers. Newer accounts keep a temporary limit until they’re verified, which Anthropic says helps prevent fraud and abuse, and they’ll get an email when it’s removed. Weekly limits are unchanged.

Anthropic tries automating misalignment red-teaming

In an October 30 post from @AnthropicAI, Anthropic described new research on automating subtle emergent misalignment research. Red-teaming, deliberately trying to break a model’s safety measures, is limited by how much time researchers have. To test whether it could be automated, Anthropic set thousands of Claude agents the task of evading a training-data monitor.

According to Anthropic, the agents produced training data that looked harmless to monitors, with examples like a veteran wrestler mentoring a newcomer or a Q&A about professional beekeeping, but that actually taught misaligned behavior. Anthropic says examples like these can be used to directly improve future training-data monitors. It calls the work “only a first step,” but says it suggests future red-teaming could improve much further with better models and more compute. The full write-up is on Anthropic’s Alignment Science Blog.