COINMAG stories about Claude, newest first, with key facts and answers to common questions.
Researchers from Princeton and Anthropic report that AI agents often forget safety boundaries after remembering harmless user requests, leading to unsafe actions in simulated high‑stakes scenarios.
Kenji SatoAnthropic used its Claude models to run autonomous agents in the space MMO EVE Frontier, achieving tutorial completion but also testing covert sabotage within teams.
Anthropic’s Claude now lets paid U.S. users link bank accounts via Plaid and query spending, balances and bills, sparking a dispute with fintech founder Anthony Pompliano.
Anthropic’s FlashProve can cryptographically prove model outputs, but generating proofs requires 70‑95× more GPU time than standard inference, limiting its immediate practicality.
Anthropic’s third‑quarter filing shows revenue growth erased by $600 million of price reductions, while compute expenses jumped 77%, turning a profit into a loss.
Social media reports indicate Anthropic may be routing traffic to an unreleased Fable 5.5, while OpenAI's GPT 6.2 Astra appears in usage data, though neither company has confirmed the releases.
Anthropic’s seven co‑founders collectively hold roughly 1.4% of the AI firm, a stake that could be worth more than $200 billion under a $2 trillion IPO valuation.