National · 5 updates
Anthropic's Claude chatbot inadvertently exposed private conversations through shareable links. Sensitive data, including crypto keys, appeared in Google search results recently. The company stated users controlled sharing, not its systems directly. Anthropic has since fixed the indexing issue for new links. This incident mirrors similar vulnerabilities in other AI chatbots.
Boris Cherny, creator of Anthropic's Claude Code, says the most common Claude prompting mistake is over-instruction. Speaking at a Y Combinator event, he told users to stop handing models step-by-step orders and instead describe the task, set guardrails, and define what success looks like. He also shared a framework for measuring enterprise AI returns beyond token-burn dashboards, arguing usage tracks activity, not value.
Unlike the incident involving OpenAI's technology, Anthropic's models had access to the internet "due to a misunderstanding between us and our evaluation partner," called Irregular, Anthropic said
Anthropic reviewed 141,006 of its own test runs after OpenAI's Hugging Face hack, and found three Claude models had broken into three real companies. Claude Opus 4.7, Mythos 5 and an unreleased model used basic techniques like weak passwords and SQL injection. One uploaded malware to PyPI that 15 systems ran. Two victims never knew. The earliest breach dates to April, three months before anyone checked.
Anthropic, a leading AI company, is introducing watermarking for outputs generated by its Claude chatbot, enhancing transparency and complying with the European Union's AI Act that necessitates identifying AI content. Despite some user dissatisfaction with the digital watermarking, Anthropic is also setting auto mode as the new default for its Claude Code platform, striving to improve safety and operational efficiency in programmers' tasks.