AI News, 16 November 2026

Monday 16 November

Perplexity Monitor API: $1 per 1,000 runs for web and page tracking

AI-generated summary Perplexity has released the Monitor API to check the web or specific pages on a schedule and send a signed webhook upon finding relevant updates. Pricing for the developer service is set at one dollar per one thousand succeeded runs for Search monitors and one dollar per one thousand fetched URLs for Page monitors.

Also covered byUnite.AIAravind Srinivas (@AravSrinivas)AI Compass
Earlier:

Apple tests daily usage caps for advanced Siri AI features in iOS 27.3 beta

Source-provided image accompanying Apple tests daily usage caps for advanced Siri AI features in iOS 27.3 beta

AI-generated summary Apple released beta versions of iOS 27.3, iPadOS 27.3, and macOS 27.3 introducing daily usage limits for advanced Siri AI and Apple Intelligence features. Higher iCloud Plus tiers will receive larger daily caps, and Logic Pro now offers a 49.99 dollar Intelligence Usage Pack for extra quota.

Also covered byRR Tech PulseZetikThe Verge
Earlier:

India’s Draft AI Rules May Require Safety Tests Before Launch for Powerful Models, Report Says

AI-generated summary According to a report by NDTV Profit citing sources, India's Ministry of Electronics and Information Technology may release a consultation paper on proposed artificial intelligence rules as early as next month. The reported framework under consideration could require powerful frontier artificial intelligence models to undergo mandatory safety tests and sandbox trials before public launch.

Vendor Statement Reconciliation in Excel with Copilot and Claude: AP Worked Example

AI-generated summary Copilot and Claude for Excel can match supplier statements and accounts payable subledgers, sort differences, and draft proposed corrections. Microsoft and Anthropic advise users to verify outputs because Copilot can make mistakes and Claude is not recommended for audit-critical calculations without verification.

Moonshine voice recognition framework survives startup shutdown as open-source project

AI-generated summary Pete Warden acquired the Moonshine speech recognition intellectual property from his defunct startup liquidation to continue the project as an open-source initiative funded by existing customers. Warden also released the Talk to Anyone mod for Kingdom Come: Deliverance II, combining the speech recognition framework with a local language model to run without cloud dependencies.

Researchers reveal “Lazarus” technique to hide malicious payloads in open‑weight AI models

Source-provided image accompanying Researchers reveal “Lazarus” technique to hide malicious payloads in open‑weight AI models

AI-generated summary Microsoft researchers introduced a technique named Lazarus that can embed a malicious command inside fewer than ten dormant neurons of an open-weight AI model. The researchers emphasize that detecting such threats would require organizations to maintain a trusted copy of the original model weights and perform diffing against downloaded versions.

Claude Prompts for Bookkeepers: 15 That Save Time

AI-generated summary Claude assists bookkeepers with writing and explaining tasks such as client emails, categorisation logic, and plain-English reports. ChatGPT is noted as suitable for quick tasks and broader comparisons across professions.

Earlier:

Investigating What Open Models Knew When They Answered Questions Incorrectly

Cloud inversion, Grand Canyon from Desert View, Arizona, Nov 29, 2013. Photo: Erin Huggins, National Park Service (public domain)

AI-generated summary Patronus.ai says it tested open models on software tasks to investigate whether incorrect answers mean the model failed to compute the answer or computed it internally and said something else. Patronus.ai says probes on activations beat model answers on rate of change tasks across five models, while service topology tasks showed no sign the answer was ever computed.