Lectura

Firecrawl announced a completed $75 million Series B round led by Smash Ventures, with Y Combinator, Altos Ventures, Nexus Venture Partners, Freestyle and Offline Ventures participating.

The company disclosed no valuation and no round terms; the amount comes from the company's own announcement as reported by SiliconANGLE. The capital is earmarked for adding third-party data sources and launching a self-service content licensing system.

Alongside the funding, Firecrawl launched a cloud service called Alexandria that combines web data with curated datasets behind a single API, aimed at data access for AI agents.

Cyera announced on September 22 that it has raised $400 million from Goldman Sachs, as an extension to the Series G round that closed in June.

That June round valued Cyera at $12 billion; the company now says only that it is worth "over $12 billion" and did not share an exact figure. The capital will fund AI-agent security features and expansion into the US federal government and international markets.

Since June, Cyera has pivoted toward AI-agent security, launching Agent Guardian and Cyera Endpoint and completing its $1 billion acquisition of Oasis Security earlier this month. The valuation language comes from the company's own announcement.

Mistral denies being hacked again, saying its investigation found no evidence of new unauthorized access.

On September 16, an account using the handle "mrwho" posted on an English-language cybercrime forum offering what it called Mistral's full source code, priced in Monero only. Mistral responded that it had "found no evidence to support this claim," but has not stated whether the listed code is genuine.

FrenchBreaches compared the 339-file tree the account shared and found at least four repository names that also appear in TeamPCP's May leak samples. No customer data has surfaced in analyzed samples, and nobody has shown files created after May 12. Deciding whether a second breach happened requires commits or still-valid credentials dated after May, which no one has verified.

The seller's account was created in September with four posts, so it could be a scam or a resale of the May dump.

In Pangram 4's own tests, AI-assisted text was misclassified as AI-Generated at rates of 4% and 7%, while the website previously showcased only the best 0.01%-order result — a gap LessWrong author Sonia Albrecht named in a September 23 post.

She says that before she contacted the company on September 17, the site displayed "99.9%+ Accuracy" directly beside "Detects AI Assistance"; the wording has since changed, though she does not know if her message caused it. In her own test, about 30% of passages in roughly 5,000 words of her writing were labeled AI-Generated.

All of this is the critic's account and personal testing; the 4% and 7% figures are not independently verified, and Pangram's paper abstract claims an overall false positive rate of 0.0041%. False positives rise on short passages, which matters for anyone using AI editing.

macOS 27 automatically downloads the Apple Intelligence local model to every Mac when it next connects to the internet, and Apple has removed the switch to turn Apple Intelligence off.

Apple's official figure is up to 14GB of local storage for M3-and-later Macs with at least 12GB of unified memory, and up to 8GB for other supported models. MacRumors measured 20.67GB on an M4 Pro Mac mini, Ars Technica recorded 22.42GB on an M3 MacBook Air, and one user's screenshot showed 30.16GB.

Even an M1 MacBook Air below the top model's hardware requirements showed about 14GB, above its 8GB cap. Apple's support documentation does not explain the discrepancy, and the measurements come from media tests and user reports, not official confirmation.

Three on-device generative AI services — Honor's YOYO Claw, Xiaomi's miclaw and StepFun's Terminal AI — completed filing with China's cyberspace regulators on September 22.

According to the official Wangxin China account, regulators announced the new filings under the Interim Measures for Generative AI Services: Honor's service is filed by Shenzhen Honor Software, Xiaomi's by Xiaomi Technology, and StepFun Terminal AI by Zhiyuan Xingchen (Shanghai), each scoped to their own brand phones.

Filing is the compliance precondition for operating such services. On-device AI had mostly appeared on cloud-model filing lists before; this batch creates a dedicated on-device category that later products can be measured against.

Samsung pushed an internal test version of its SmartThings software to consumers, and some fridges stopped working entirely after installing it. The company has halted distribution and publicly apologized.

The failures appeared in South Korea on the afternoon of September 22, days before the Chuseok holiday. Most affected units are from the Bespoke AI four-door series launched in 2024 or later; dozens of complaints appeared on Samsung's official forum, and some users said holiday food had to be thrown away.

The fridges require an on-site engineer visit, and one user was told service might not arrive until after the holiday. Samsung has not disclosed how many units were affected or announced compensation for spoiled food. For readers, this is a concrete case: the update channel of a connected AI appliance can deliver a vendor's test software straight into devices in service.

Tesla announced on X on September 23 that Grok Bot is live in North American vehicles, letting drivers check email, manage schedules, shop online and place orders hands-free.

The agent runs inside the existing Grok interface rather than a separate app; access to third-party services like Gmail and Google Calendar requires setting up Connectors. Early testers showed a workflow ordering Starbucks via voice while driving with FSD, but that remains vendor and tester claims, not independently verified.

Availability is currently limited to SuperGrok Heavy subscribers, with other tiers to follow; SMS support is still in development, so Grok can only place calls for now.

Snowflake announced on September 22 that Claude Opus 5.5 is available in public preview on Cortex AI, across CoCo, CoWork, Cortex Agents, AI Functions and Cortex Inference.

All calls stay within Snowflake's security and governance perimeter, and developers can also reach the model through the Anthropic SDK via Cortex Inference.

The coding, accuracy and price-cut figures quoted in the post come from Anthropic's own announcement and are vendor-reported; this is a public preview, not general availability.

Databricks announced on September 22 that the Genie One MCP server is now generally available to all users, ending its preview phase.

The interface lets any MCP-compatible agent — Claude, ChatGPT, Cursor or internal tools — reach Genie One's structured and unstructured data answers through a single entry point, with semantics constrained by Genie Ontology to reduce conflicting answers across separately configured agents.

Governance and audit logging run through Unity Gateway, which is what determines whether enterprises can approve it. Note this is a vendor announcement with no independent evaluation; the customer quote was chosen by the vendor, and real answer quality and adoption remain to be seen.

Teradata announced on September 22 that it is expanding its Tera AI assistant with three additions — a Context Engine, Tera Harness and Agent Skills — available in the fourth quarter of 2026.

The Context Engine brings together metadata, lineage, business definitions and access policies without moving data, so agents interpret requests in organizational context. Tera Harness handles execution: it selects tools and models, tracks multi-step progress, can pause for human approval before sensitive actions, and resumes after infrastructure failures. Organizations can connect their own tools via the Model Context Protocol.

The announcement also cites company-reported benchmarks: using the same Opus 5 model, Tera used 73% fewer tokens than Claude Code, and Teradata reported 53% lower cost per reliably solved task than Snowflake's Cortex Code. These are Teradata's own tests — the company itself cautions they do not reflect all enterprise workloads — and real-world value can only be judged after the Q4 release.

In a product note dated September 22, OpenAI said the GPT-6 family delivers higher prompt cache hit rates by default, with cache discounts for eligible shared prefixes reused within a 30-minute window and discounts of up to 90% on cached input tokens.

New tooling includes a caching dashboard, a cache-miss diagnostics tool, and explicit cache breakpoints. Developers can now change reasoning effort between responses without breaking cache, and prewarm known context to cut first-token latency.

Customer figures in the post are self-reported: GitHub Copilot says the share of prompt tokens needing fresh processing fell by more than 50% over recent months, Manus says its cache hit rate rose from roughly 85% to above 90%, and Strawberry Browser reports a 36% inference cost reduction. These are vendor and customer claims, not independently verified.