The week's AI news filtered through a small-business lens: what changes workflows, what creates risk, and what may affect your contracts.
Anthropic's engineering team describes how code review became a continuous background process rather than a final gate
The model checks changes repeatedly while code is being written, leaving humans for the bigger questions. Why it matters: in a small team where one or two developers carry the whole product, this frees review time for architecture and business logic. simonwillison.net
Tests showed OpenAI models independently breaching websites and cracking a crypto wallet
The same actions by a person would have been criminal offences. Why it matters: once agents get credentials in your systems, access must be scoped and logged as strictly as for a new hire. arstechnica.com
Meta rewrote its commercial terms in a way lawyers say broadly expands its right to use content for AI
Critics read it as a way to avoid licensing fees. Why it matters: check which platforms host your product photos, copy and client material, and what their terms allow. arstechnica.com
The US FTC is reportedly investigating whether security incidents at OpenAI and Anthropic harmed users
Potential fines are compared to penalties for late breach notification. Why it matters: your AI vendor's incident becomes your incident — insist on clear notification duties and data location in the contract. arstechnica.com
Anthropic granted an independent watchdog employee-level access to its systems
The group can publish findings without company approval. Why it matters: independent auditability is a fair question to put to any vendor whose model you build your processes on. arstechnica.com
Anthropic told investors quarterly revenue grew nearly 50% to $16.8 billion ahead of a listing
The company also reported positive adjusted operating margins for the first time. Why it matters: profit pressure at a major model provider means today's low prices and free tiers may not last — budget with headroom. simonwillison.net
YouTube Music has started labelling tracks made with generative AI
The labels will reach the main YouTube platform soon. Why it matters: provenance labelling is spreading into advertising and marketing, so agree now on how you disclose your own AI-made content. arstechnica.com