Amazon reportedly triggered the US shutdown of Fable 5 and Mythos 5
A day after Anthropic pulled Fable 5 and Mythos 5 offline on a Commerce Department directive, reporting pointed at Amazon, Anthropic's own backer, as having tipped off Washington (under an effort referred to as "Project Glasswing") over a jailbreak risk. The ban blocks visa holders, API customers and frontier researchers alike, with even Andrej Karpathy, an EB-1 holder, unable to use Fable, and the industry reaction split between anger at the government and anger at Anthropic, amid talk of "sovereign AI" and warnings that compliance is near-impossible. This is a fresh twist on the launch-week controversy logged on 13 June. The Wall Street Journal first reported that Amazon CEO Andy Jassy's talks with US officials triggered the crackdown.
ElevenLabs gives its AI voices a face
ElevenLabs launched ElevenCreative, which pairs its studio-grade synthetic voices with a talking-head avatar: feed it a script, pick a voice and a face, and it renders a finished talking-head video in one place.
HashiCorp's Terraform MCP Server reaches general availability
HashiCorp announced general availability of the Terraform MCP Server, an open-source MCP server that lets AI agents integrate with Terraform Registry APIs, pitched at lifting infrastructure-team productivity by taking on rote tasks.
InfoQ →
Agents · Infrastructure
HashiCorp
SwirlAI argues against monitoring AI systems like web services
A SwirlAI essay contends that LLM and agent systems break the assumptions behind web-service monitoring (uptime, error rates and latency percentiles) and should instead be measured against five questions: is it fast, can it scale, is it correct, does it hold up, and how does the agent behave. It makes the case for treating latency as positions on the prefill-and-decode timeline (such as time-to-first-token), for making cost per successful task the connective number rather than cost per request, and for building quality signals such as LLM-as-judge scores calibrated against human labels, which the serving stack never emits on its own.
"MCP solved tool calling, A2A solved coordination, what solves transport?"
A VentureBeat piece argued the agent stack still lacks a settled transport layer: MCP standardised how agents call tools and A2A how they coordinate, but moving messages reliably between agents and services across networks remains the unsolved problem, with implications for how production multi-agent systems are wired together.