Anthropic can’t reliably control its AI agents — it’s cutting internal evals off from the live internet
The control problem, stated plainly by its biggest practitioner: Anthropic says it can't reliably control its AI agents, and is disconnecting internal evaluations from the live internet to contain them. When the safety company air-gaps its own evals, the agent-control debate moves from theoretical to operational.
The false homicide tip, day two: it went to a spam folder — and Anthropic self-reported
The Philadelphia story develops: per FOX 29, the AI-generated tip went to a spam folder and police never acted on it. Anthropic notified the department and published its own incident report — a notable self-disclosure that sets a precedent for how model vendors handle real-world misfires.
Nvidia in talks to acquire ‘open’ model startup Reflection AI
Consolidation watch: Nvidia is reportedly in talks to acquire Reflection AI — maker of Beam, the 501B open-weight model we covered this week. The open-weights upstarts are now acquisition targets for the GPU giant.
The maker of non-text AI model Jev valued at $7.5B just weeks after launch
Beyond language: the startup behind Jev — a non-text AI model — is already valued at $7.5 billion, weeks after launch. The market is pricing modalities beyond words.
Apple discloses deal to hire team and license tech from podcast startup Huxe
Apple continues its quiet acqui-hire strategy: it's hiring the Huxe team and licensing its personalized podcast technology — the personalization layer of Apple Podcasts just got a specialist infusion.
Here are the top AI agents that can live in your text messages
The agent interface is the inbox: TechCrunch rounds up the AI agents that now live inside your text messages — no app install required.