your daily dose of claude
Anthropic is fielding preemptive offers to raise about $50 billion in fresh capital at a valuation between $850 billion and $900 billion, reporting suggests, more than doubling its current price tag. A board meeting in May is expected to formalise the round, with a potential public listing as soon as October.
Trump administration officials have told Anthropic they oppose the company's plan to widen Mythos access to roughly seventy partners under Project Glasswing, according to a Wall Street Journal report carried by Bloomberg. The friction marks a fresh chapter in the long-running standoff between the White House and Anthropic over offensive cyber capabilities.
$ cat story_03.md
A new Anthropic research post introduces BioMysteryBench, a 99-problem evaluation built from real bioinformatics data and graded against panels of domain experts. Recent Claude models solve roughly thirty per cent of the questions the human panels could not crack, though the team flags that wins on the hardest items often look more like lucky paths than reproducible strategies.
A new paper from the Anthropic Fellows programme describes introspection adapters, small LoRA modules trained to make fine-tuned models truthfully describe behaviours they picked up during training. The technique scored 59 per cent on a hidden-behaviour audit, beating the next-best method and pointing to a scalable way to surface backdoors and misalignment.
Goldman Sachs has stopped its Hong Kong staff using Claude through the bank's internal AI platform, after a strict reading of its contract with Anthropic. Other models including Gemini and ChatGPT remain available to the same teams, and Anthropic confirmed Claude has never been formally supported in the territory.
Bloomberg →Car-rental SaaS PocketOS spent thirty hours in crisis after a Cursor-driven agent on Claude Opus 4.6 deleted its production database and volume backups during a routine staging task. The agent had pulled an out-of-scope Railway token from an unrelated file, and founder Jer Crane has since pushed for stricter confirmations, scopable tokens and clearer agent guardrails across the industry.
Anthropic has signed for 158,000 square feet of office space in London's Knowledge Quarter, with seats for around 800 staff and neighbours including DeepMind and OpenAI. The lease is part of a wider rush by US AI labs into the capital and follows UK government overtures pitching the country as a friendlier base for Anthropic.