AI DIGEST: 8–14 JUNE 2026

Five things from the second week of June, picked out of everything we published on the daily feed.

Claude Fable 5 Lands

Anthropic released Claude Fable 5 for general use, alongside Claude Mythos 5 for selected cyberdefenders and infrastructure providers, with the headline gains in software engineering, research and security work. Ethan Mollick's account of using it is the more informative read: he describes commissioning whole projects (coordinating agents, running research, writing and testing code) rather than directing each step. NVIDIA also shipped Nemotron 3 Ultra, a 550B hybrid Mamba-Transformer with a 1M-token context, so the frontier moved on both the closed and open sides in the same week. If your evaluation harness still measures single-turn answers, it is now measuring the wrong thing.

Invisible Safeguards, and a Fast Reversal

Within days of the launch, researchers reported that Fable 5 was quietly degrading or refusing work related to training competing models, debugging AI code and optimising neural architectures, described as safeguards that users could not see firing. Anthropic walked the policy back and committed to making the interventions visible, and Nathan Lambert used the episode to argue that uneven safety policy erodes the trust vendors depend on. The lesson for buyers is procedural rather than political: silent quality variation is a supply chain risk, and the only defence is your own regression suite, run on your own tasks, on a schedule.

Multimodal AI That Fits on a Laptop

Google released Gemma 4 12B under Apache 2.0: native vision and audio, fed straight into a decoder-only transformer with no separate encoder, running on 16GB-class machines. The quantisation-aware checkpoints followed immediately, cutting memory by around 72% with near-original quality and squeezing the smallest text-only variant under 1 GB. That combination of multimodal, offline, permissively licensed and commodity hardware is what makes on-premise viable for workloads where data cannot leave the building. It is also a reminder that "which model" and "where it runs" are increasingly the same decision.

Coding Agents Are Bad Teammates

Stanford's CooperBench found that pairing coding agents on real software tasks often makes results worse than a single agent. The bottleneck is coordination rather than coding ability, and adding a messaging channel between them barely helps. This runs against the reflex to solve a hard task by adding more agents, and it points at where the engineering effort actually belongs: explicit task decomposition, clear ownership of state, and a human or a deterministic orchestrator holding the plan. As one widely shared post put it this week, the model is no longer the bottleneck, the workflow around it is.

Agents Are Growing Up as Infrastructure

Three pieces of plumbing landed together. Uber described giving each agent a verifiable cryptographic identity with short-lived per-hop tokens that preserve the full delegation chain, solving the "who authorised this action, on whose behalf" problem that audit teams ask about first. LangChain shipped hardware-virtualised microVMs so agents can run untrusted code away from production, and NVIDIA released SkillSpector to scan agent skills for malicious patterns before installation. Identity, isolation and supply chain scanning are exactly the controls that already exist for services and packages. Agents are simply catching up to them.