The Signal — September 22, 2026

Amazon blocked Meta's Muse from shopping on its site after raising security and authorization concerns. Artificial Analysis scored Grok 4.7 twelve points below xAI's launch table, while NVIDIA researchers tested an AI-designed agent harness against acceptance rules frozen before the search began.

Amazon threw Meta's shopping agent out the same week Shopify planned a checkout

Since late September 20, anyone using Meta's Muse assistant to shop on Amazon.com has gotten a popup instead of a purchase: "Continued access by an unauthorized AI agent violates Amazon's Conditions of Use, to which our customers have agreed." Amazon told GeekWire it was never told Muse would be shopping its store, never authorized it, and had already asked Meta to carve Amazon out before resorting to a block. Amazon says Muse captures and stores customer credentials, scrapes account data, and browses without identifying itself as a bot.

The block cuts across an existing partnership. Amazon products have been purchasable inside Facebook and Instagram since 2023, and Meta signed a multibillion-dollar deal in April to run agentic AI workloads on Amazon's cloud. Amazon also earns more than $68 billion a year in ad revenue from people looking at its pages, something an agent doing the shopping does not do. It spent the past year keeping outside agents off the site, suing Perplexity over its Comet browser and blocking Google's and OpenAI's shopping agents. It won an injunction against Perplexity in March and lost it on August 4, when the Ninth Circuit held that the user, not the AI company, is the party accessing Amazon's computers under federal anti-hacking law. The current block relies on terms of service and a popup rather than another lawsuit.

Shopify spent the same week going the other direction, with the Wall Street Journal reporting plans to let Muse complete purchases through Shop Pay. "Agentic commerce really excites us," Shopify's Rohit Mishra told American Banker. "The agents don't have the same challenges as humans." Reporting differs on how much of that is live versus announced. Stripe's Link is the currently documented payment path, so Shop Pay checkout should be treated as coming rather than shipped. In a separate finding, macOS researcher Patrick Wardle published a zero-day in the Muse Mac app that let any local app or terminal command redirect the agent's transcription endpoint and take the token authenticating a user's account. "We can manipulate the agent and leverage its privileges to do whatever we want," he told Ars Technica. Meta says it has issued a hotfix. Wardle's finding supports the general security concern but does not verify Amazon's account of what Muse does with credentials. Meta did not respond to GeekWire's request for comment.

Sources: GeekWire · TechCrunch · Ars Technica · American Banker

Xiaomi's open-weight model tied Grok 4.7 on an outside index, at a fifth of the input price

xAI released Grok 4.7 on September 21, built on a larger base model than 4.6 and trained with a longer reinforcement learning run weighted toward problems that take hours rather than minutes. It carries a 500K-token context window at $2 per million input tokens and $6 per million output. Xiaomi shipped MiMo-V2.6 Pro and Flash the same day. The open-weight, MIT-licensed mixture-of-experts model has a trillion total parameters and 42 billion active, a million-token context window, and text, image, speech and video input. On Artificial Analysis' Intelligence Index, a composite of ten evaluations run by an outside party rather than either vendor, both score 46. Artificial Analysis prices the Xiaomi model at $0.435 in and $0.87 out, a fifth of Grok's input price and a seventh of its output.

On Terminal-Bench 4.0, xAI's own launch table reports Grok 4.7 at 38.0 percent. Artificial Analysis, running the same named benchmark, measured 26 percent, a gap wide enough that the cheaper DeepSeek V4.1 Flash edges past Grok at 27. Xiaomi's own table puts MiMo-V2.6 Pro at 34.9 on the same test. The discrepancy does not by itself show that anyone is falsifying a score. Effort settings, harness configuration and benchmark versions differ between vendor tables and outside evaluations, and those choices can move scores by double digits. The shared benchmark name therefore does not guarantee a shared measurement setup. Artificial Analysis is independent, but it is one evaluator running one configuration, and its index is versioned and moves. The 46-to-46 result applies to that index on that day; it does not establish that the models are interchangeable.

Sources: xAI · Artificial Analysis (Grok 4.7) · Artificial Analysis (MiMo-V2.6-Pro) · Hugging Face · The Decoder

NVIDIA let an AI audit a coding agent's plumbing and it cut a third from the bill

Yesterday's TypeSafe story looked at the decisions consuming agent budgets. Researchers at NVIDIA, NTU and MIT approached the cost question through the harness around the agent. They posted SoL-Pi to arXiv on September 17; the code and project page followed on the 21st. A research AI watched the open-source Pi coding agent work, proposed changes to its harness, and tested them across 152 directions, 535 executable environments, more than 3,000 runs and over 60,000 agent-environment interactions. Four efficiency mechanisms survived.

The four mechanisms address ordinary sources of waste. Action Fusion merges the file edit and the command that tests it into one tool call, skipping a model round trip. ObservationPack stops replaying large tool outputs into every subsequent request, archiving them locally behind a handle and a short excerpt, with exact pages retrievable on demand. A third mechanism compacts context at subtask boundaries, but only when projected savings exceed the cost of rewriting the prompt cache. The fourth sends long build logs to a cheaper model, then has a deterministic verifier check the returned summary's schema, source hash, exit status and exact quotes before it is allowed into context. On the held-out EdgeBench suite the stack cut token traffic 49 percent and API cost 33 percent on GPT-5.6 Sol, and 44.7 percent of tokens on Opus 5.

Acceptance rules were frozen before the search started, the optimizer could not change them, and held-out results never fed back into the search. Those controls limit the optimizer's ability to overfit the rules it is scored on, although the apparatus remains self-administered and the paper is a preprint. The efficiency gain also carries a capability cost. The harness keeps about 94 percent of Pi's average score. On Terminal-Bench 4 it solved 15 tasks where Codex and Pi both solved 18, for 26 percent less money, and in a swarm test 20 SoL-Pi workers reached fewer cycles for less money than 20 Pi workers while a single Codex agent came in cheaper than either arrangement. The team calls cross-model transfer preliminary, noting the mechanisms fire less often on Opus 5 because the search only ever watched GPT-5.6 Sol.

Sources: arXiv · NVIDIA Labs · GitHub · MarkTechPost

On the Editor's Desk

OpenAI published a proposal asking the United States to lead an international standards effort for frontier AI, covering how to evaluate recursive self-improvement, when automated research should trigger human review, and how incidents get classified and reported. We have run OpenAI's alignment posture repeatedly this month, and the document asks for standards rather than establishing any. California also signed seven data center laws covering electricity, water and land use, including a new utilities rate classification and the end of blanket environmental exemptions. We covered the state's AI executive order three days ago and spent a weekend feature on who pays for the buildout, so this would have been our third pass at California in a week.

Also held: an FDA discussion paper on regulating generative-AI medical devices that turned out to be from August 18, and which the FDA itself says is neither guidance nor a policy change. AWS released an open-source agent harness the same week as the NVIDIA work, with a 28 percent cost-savings claim from its own six-benchmark test and a paper still pending; running both would have made two-thirds of this edition about token bills.