Squared: The Week Agents Got Real (and Anthropic Got Caught)

Anthropic shipped Claude Fable 5 with invisible guardrails that quietly degraded the model for researchers and competitors using it to build rival systems. After a Wired scoop, the company reversed course and promised transparency about when restrictions kick in. Separately, Fable 5 now flatly refuses cybersecurity, biology and chemistry queries.

This is a vendor-dependency risk, not a model-quality story. If a provider can silently change behaviour to protect its own commercial interests, your evaluations and your production outputs are sitting on shifting ground. Revisit two things: are you benchmarking models on a fixed snapshot, and do you have a fallback provider? Don't build a critical workflow on a single closed model whose owner can throttle it without telling you.

A German court ruled against Google's AI Overviews, and Ars Technica reports the decision could ripple across the AI search industry. The reasoning — that AI summaries aren't a necessary feature of search — opens the door to more challenges in the EU.

If you've quietly bet on AI-generated answers replacing search-and-click as a traffic or product strategy, the regulatory floor is less solid than the vendors imply — especially in Europe. This won't kill AI search, but it raises the odds of region-specific restrictions. Anyone with EU exposure should treat AI-summary features as legally provisional, not permanent.

OpenAI announced plans to acquire Ona to give Codex secure, persistent cloud environments — the plumbing that lets AI agents run long, multi-step tasks across enterprise workflows. It lands alongside DeepMind funding multi-agent safety research and warning about what happens when millions of agents interact.

The direction is clear: agents that act autonomously over hours, not chatbots that answer questions. MIT Technology Review reports agent adoption could surge up to 300% in two years. The leadership question isn't whether to adopt — it's governance. Before you let an agent run unsupervised in your environment, decide what it's allowed to touch, how you audit it, and who's accountable when it goes wrong. DeepMind funding safety research tells you the experts don't have the answers yet either.

DeepMind released DiffusionGemma (4x faster text generation) and Gemma 4 12B; NVIDIA optimised both for local hardware; Gemini 3.5 Live Translate brought near real-time voice translation to Meet and Translate; India's Avataar priced video generation at $0.005 per second.

This is the steady commoditisation of capability — genuinely useful, rarely urgent. Faster local models lower the cost of running AI on your own hardware, which matters for data-sensitive workloads. But none of it requires action this week. File it under 'the cost of doing this keeps falling' and revisit your build-vs-buy maths quarterly, not weekly.