Anthropic hid a traitor in a 4-Claude game team

🤖 Claude Code Blocks Game Cloning Anthropic added 2 sentences to the system prompt for Claude Code on Opus 5.5. The rule blocks recreating a specific existing game, even when it is named directly or presented as original. The change reportedly reached servers around October 27 under the flag tengu_heron_brook. Users can still make original games, genre projects, or work based on their own designs and assets. Anthropic’s Addy Osmani confirmed the prompt was changed, but not the exact wording. The stated goal is to stop Claude from reproducing other people’s games, which violates its usage rules. Similar changes elsewhere in Claude remain unknown. 📊@tech

Cells contain tiny hollow protein barrels called vaults. They've been known since 1986, and scientists still haven't pinned down what they do. Anthropic's biology team published a story yesterday about using Claude to investigate their relatives in bacteria and the viruses that infect them. The team made vault-like particles in E. coli and mapped one phage-derived vault's structure. They still don't know its biological function. The best part of the write-up isn't a speed claim. It's the sequence of disagreements: Claude initially argued against a faster method for isolating the particles. The researchers tried it anyway, with Claude helping refine it. Later, Claude suggested putting raw broken-cell contents straight onto an expensive column. The humans were reluctant. That worked too. The

Anthropic loosens Claude's game cloning ban https:// fed.brid.gy/r/https://www.toms hardware.com/tech-industry/artificial-intelligence/anthropic-loosens-game-restrictions-claude-will-now-offer-to-build-original-game-in-same-genre-if-asked-to-copy

ChatGPT's Temporary Chat and Claude's incognito chats keep a conversation out of your history and memory, and neither gets used for training. But OpenAI says it may keep a copy for up to 30 days for safety, and Anthropic keeps incognito chats for 30 days by default. You're also signed in to your account the whole time, and on Claude's Team and Enterprise plans, incognito chats are included in data exports for account owners. So they hide the chat from your history, not from the company. The most private option is running a model locally. Your prompts never leave your computer and there's no account involved. Tools like Ollama and LM Studio make the setup pretty easy, and open models like Qwen and DeepSeek are free to download. The catch is hardware: small models run on a decent laptop, but

Anthropic reported on November 5 that Claude helped determine the first three-dimensional structure of a vault particle: a hollow protein shell built from a protein encoded by a bacteriophage, a virus that infects bacteria. The work is posted as a preprint on bioRxiv, so it has not been peer-reviewed yet. According to Anthropic, the project ran for under four months. Its science-focused version of Claude scanned about 15 billion computer-predicted proteins to hunt for vault-like structures, and the team then worked out this structure experimentally. Vault particles are strange, hollow shells whose biological role is still poorly understood, which makes this an interesting case of AI doing more than speeding up a known pipeline: it helped pick the target in the first place. A few things to

🎭 Someone spent $20,000 so AIs could lie to each other at Werewolf Podcaster Dwarkesh Patel paid about $20,000 in API credits to make frontier models play Blood on the Clocktower — a social-deduction game like Werewolf — against each other for 386 matches. One model plays the whole good team; another plays the whole evil team. OpenAI's GPT-6 Astra leads the board at about 1,859 Elo. Against Claude Opus 4.6 it wins roughly 92% of the time. Dwarkesh says the edge is mostly catching liars, not telling better ones. Full rankings and replays are on clocktowerbench.com. You can also play the AIs yourself. @aipost 🏴