Edition, Tue, 22 Sep 2026.

Jev introduces a new class of "decision models" that trade text for typed probabilistic outputs

TypeSafe AI unveiled Jev, the first example of what the company calls a "System One model" — an entirely different architecture from traditional large language models.

TypeSafe AI unveiled Jev, the first example of what the company calls a "System One model" — an entirely different architecture from traditional large language models. Instead of generating free-form text, Jev takes unstructured input and returns structured probabilistic decisions: confidence scores for yes/no answers, probability distributions across choices, or floating-point scores for numeric scales. The model processes questions in parallel, charges only for input tokens at $0.042 per million, and is already sparking an open-source ecosystem with projects like Kev attempting to replicate Jev on open-weight models such as Qwen 3.5.

The category promises dramatically cheaper classification, spam detection, and reranking tasks. But Simon Willison, who documented the technology, warns that the lack of explainability — Jev returns only a number with no reasoning — makes it risky for sensitive decisions like hiring or lending, and early tests revealed troubling societal bias in its city ratings.

Meta patches zero-day vulnerability in its Muse AI agent that grants attackers full account control

Security researcher Patrick Wardle discovered a zero-day vulnerability in Meta's new Muse macOS app that allowed any locally running application or terminal command to redirect the agent's transcription endpoint to an attacker's server.

Security researcher Patrick Wardle discovered a zero-day vulnerability in Meta's new Muse macOS app that allowed any locally running application or terminal command to redirect the agent's transcription endpoint to an attacker's server. By exploiting an undocumented setting, Wardle was able to take complete control of the Muse account — including taking photos, writing files to disk, and accessing connected services like WhatsApp, email, and social media accounts. Meta issued a hotfix hours after the Ars Technica report went live.

The exploit arrives just weeks after Meta's launch of Muse, which has already seen estimated downloads outpace ChatGPT's debut in its first 12 days. Amazon has since blocked Muse from accessing its e-commerce platform, claiming Meta never obtained permission. Meta's David Singleton characterized the vulnerability as a local privilege escalation rather than a remote exploit, arguing the practical risk to end users was low since it required malicious code already running on the victim's machine under their account.

OpenAI creates independent math advisory group after its AI resolves more than 100 open problems

OpenAI announced the creation of an independent advisory group hosted at the Institute for Advanced Study in Princeton, designed to give the mathematical community a voice in the company's AI-driven math research.

OpenAI announced the creation of an independent advisory group hosted at the Institute for Advanced Study in Princeton, designed to give the mathematical community a voice in the company's AI-driven math research. The move comes after OpenAI's internal models solved more than 100 open problems across major areas of mathematics, including the abrupt publication of a solution to the Navier-Stokes Millennium Prize problem in early September that drew sharp criticism from the mathematical community.

The group will assess the significance of new results and coordinate their release, but OpenAI made clear the advisory group will not control the pace or direction of internal research. Nine prominent mathematicians were named as initial members, including only one — Camillo De Lellis of the Institute for Advanced Study — who also signed an open letter from 25 Fields Medalists arguing that AI labs are threatening mathematicians' intellectual work by racing to publish solutions at a frenzied pace.

xAI's Grok 4.7 leads the Harvey legal agent benchmark at $2 per million tokens

xAI claims that Grok 4.7 achieved a 19.6 percent pass rate on the Harvey legal agent benchmark, outperforming both Anthropic's Claude Fable 5.1 at 6.7 percent and OpenAI's GPT-5.6 Sol at 2.5 percent.

xAI claims that Grok 4.7 achieved a 19.6 percent pass rate on the Harvey legal agent benchmark, outperforming both Anthropic's Claude Fable 5.1 at 6.7 percent and OpenAI's GPT-5.6 Sol at 2.5 percent. The model is priced at $2 per million input tokens, significantly undercutting the $10 per million rate of comparable models from Anthropic and OpenAI. Independent verification by Artificial Analysis confirmed xAI's reported scores on several other benchmarks, though it found a roughly 12-percentage-point gap on Terminal-Bench 4.0 that remains unexplained.

The Harvey benchmark measures performance on real-world legal tasks, making it a notable indicator of an AI model's practical utility in professional settings. Grok 4.7's dominance on this benchmark positions xAI as a strong contender in the emerging market for legal AI agents, even as independent reviewers flagged the model's high verbosity as a potential hidden cost that could narrow its per-task price advantage.

Apple faces $250 million class-action payout as eligible iPhone owners begin filing Siri AI settlement claims

Apple must now pay out up to $250 million to settle a class-action lawsuit alleging it failed to deliver on its AI-upgraded Siri promises.

Apple must now pay out up to $250 million to settle a class-action lawsuit alleging it failed to deliver on its AI-upgraded Siri promises. The settlement opens a claim window through December 21, 2026, for U.S. customers who purchased an iPhone 15 Pro, iPhone 15 Pro Max, or any iPhone 16 model between June 10, 2024, and March 29, 2025. Eligible owners can receive an estimated $25 per device, potentially rising to $95 depending on total claim volume.

The lawsuit centered on Apple's marketing of iPhones 15 Pro and iPhone 16 as "built for Apple Intelligence" at launch, despite a significantly more limited set of features than those previewed at WWDC 2024. Apple delayed the upgraded Siri rollout for more than a year, and only recently shipped the AI-upgraded version with iOS 27. Apple has consistently denied wrongdoing, maintaining that feature availability timelines are subject to change.

Xapien lands on the ServiceNow AI Platform

The AI due-diligence startup's checks will now run natively inside ServiceNow workflows, with free trials in the ServiceNow Store.

The AI due-diligence startup's checks will now run natively inside ServiceNow workflows, with free trials in the ServiceNow Store.

Xapien, the AI-native due diligence startup, is now available as a native application on the ServiceNow AI Platform, and the app is live in the ServiceNow Store with free trials for ServiceNow customers. It delivers fully-sourced, human-grade due diligence on any person or organization in minutes, embedding research directly into third-party risk management and onboarding workflows so risk, compliance, procurement, and business development teams can vet counterparties without leaving the system.

The move lands less than two weeks after Xapien's $56 million investment round from Spectrum Equity and YFM Equity Partners and its U.S. expansion plans. Xapien's Chief Strategy & Partnerships Officer Zach Rothstein said the partnership puts compliance-grade background checks inside the workflows where clients already run their risk and onboarding operations, while ServiceNow Group VP of Global Technology Partnerships Alix Douglas called it ecosystem innovation that turns insight into action.

Cantor Fitzgerald lifts ServiceNow target to $174

A big price-target hike, tied to the company's security-driven agentic AI push, helped push NOW up 1.6% on Monday.

A big price-target hike, tied to the company's security-driven agentic AI push, helped push NOW up 1.6% on Monday.

Cantor Fitzgerald analyst Thomas Blakey raised his price target on ServiceNow from $141 to $174 while maintaining his Overweight rating, and shares of the enterprise software company rose 1.6% on Monday. In his note, Blakey wrote that ServiceNow is attracting more investor interest because of its emphasis on security-driven, agentic AI platforms, and that checks with clients showed they are actively upgrading to higher tiers that generate more revenue for the company.

With that tailwind, Blakey believes ServiceNow can reach management's goal of 30% of total annual contract value coming from AI by 2030. The stock has recovered from the broad software sell-off earlier this year and remains well down from its peaks, which the Motley Fool says makes it a bargain relative to its AI-enhanced growth potential.