This edition leans on security and on how AI tools actually make their way into companies. Seven items with direct consequences for how a smaller business adopts AI.
Google confirmed that Gemini broke into three real companies' systems during controlled tests
In the trials the model guessed passwords and dug access credentials out of public code repositories. Why it matters: keys and passwords forgotten in old repos are now findable at machine speed — auditing and rotating credentials is the cheapest security work you can do this month. simonwillison.net
Researchers used Claude to reach an OpenAI employee account and sensitive code data
The attack went after a person and their access rights rather than a technical flaw. Why it matters: if that works against a major AI lab, then in a 20-person Estonian company two-factor authentication and least-privilege access are basic hygiene, not an optional extra. arstechnica.com
Developers in the Rust ecosystem are being targeted with fake job and partnership offers
A video call is arranged, after which the target is nudged into installing a "missing codec" or running a command pasted from the chat. Why it matters: the same trick works on any developer, accountant or manager — agree as a team that no file or command received during a call gets executed without verification over a second channel. simonwillison.net
Claude Code now falls back to AGENTS.md when no CLAUDE.md is present
That moves the tool toward a shared instruction-file format already supported by other coding agents. Why it matters: project rules written once now work across several tools, which cuts vendor lock-in and speeds up onboarding for both new hires and new agents. simonwillison.net
An engineer describes a large company where specs and code are all agent-generated and nobody reviews them
Management pushes for more output on the grounds that writing code is no longer the bottleneck. Why it matters: a clear warning about turning AI into a volume metric without review or ownership — measure delivered business value, not lines produced. simonwillison.net
A US government website briefly ran an open Chinese model that the FBI had flagged as malicious
The tool slipped onto the site without anyone checking where it came from. Why it matters: AI components arrive through subcontractors and libraries — keep a simple inventory of which models and services get to see your data. arstechnica.com
A practical rule: use the model as a copyeditor, never as the writer
Let it check facts, spelling and clarity, but refuse to adopt the phrasings it proposes. Why it matters: it keeps your proposals, client emails and website copy in your own voice and avoids the recognisable "AI smell" that erodes customer trust. simonwillison.net