Last week’s review was about the receipts. This week was about the keys. Almost every story that mattered came down to the same question: when software acts on your behalf, what exactly is it allowed to hold?
Defense found it first, this time
The Monday edition opened with companies hedging against an AI crash. By Wednesday the mood had shifted from money to trust. Researchers using frontier models turned up two critical remote-code-execution bugs in PHP that, by the companies’ accounts, nobody exploited before the patches shipped. The same edition carried the counterweight: a poisoned LiteLLM release on PyPI that installed a credential stealer, and a Railway security disclosure. AI is good at finding old holes, and the AI toolchain is a fresh place to lose secrets. Both facts held at once.
Agents get wallets
Thursday’s edition was the clearest statement of the week’s theme. Anthropic’s Claude Duet arrived as a waitlisted agent app, and 1Password announced a way for agents to log in without seeing your password. Giving agents wallets and logins is the product direction. Delegated credentials that the agent can’t read are the safety design. Whether that design holds is the question for the next few months.
Offline by choice
Friday’s news landed hardest. OpenAI said it will stop training its frontier models on the live internet, citing significant risks, weeks after an engineer said that approach wouldn’t work. It is the same instinct as the credential vaults, applied to the model itself: keep it away from the thing it can be hurt by. The week’s funding talk of a $30B raise at $1.25T ahead of an IPO is the reason such caution will be scrutinized.
What mattered less than it seemed
The Anthropic IPO saga produced a lot of heat, including a safety bill Anthropic drafted for itself and a leaked filing, but no new fact that changed the picture. The DeepSeek story, a federal ban and a reported $15B raise within days of each other, said more about the limits of export-style restrictions than about DeepSeek. And the Saturday and Sunday editions, covering Kimi K3.1 Pro and an election study of Grok and GPT, were reminders that the frontier is now crowded enough that who checks the model matters more than which model is ahead.
The pattern from both of the last two Sundays: AI’s story has stopped being about capability and become about custody, of money, of credentials, of the truth. Expect the next fights to be about the same thing.
