Microsoft Expands Copilot as Databricks Patches Security Flaw

Microsoft is expanding Copilot across the workday with a new Home page for Chat and Cowork, an Auto feature that selects models and effort, and Work IQ, which searches Microsoft 365’s more than 19.5 exabytes of data. Copilot can create Word, Excel and PowerPoint files, while Autopilot agents receive Microsoft Entra Agent IDs and managers. Autopilot is rolling out to Frontier customers; Code remains weeks from general availability, and the Today command center is in private preview. Microsoft also reports that more than 5,000 Lilly employees use Cowork and introduces an Agent 365 Business Analytics tool to measure completed work.

Several other companies are extending AI into everyday workflows. Linq raised a Series B led by a16z for messaging tools that connect businesses and AI agents to channels including Apple Messages for Business, RCS and WhatsApp; its API serves more than 1,000 teams. DoorDash shoppers can browse food, build a cart and check out in one message thread. OpenAI demonstrated Codex Security surfacing only P0 and P1 vulnerabilities in GitHub pull requests, with no pricing or rollout date listed. Lovable says it removed its agents’ 30-minute limit using checkpoints that restore in under a second; rewriting its development server in Rust cut preview startup from 40 seconds to three and reduced memory use 6.5-fold.

Databricks is addressing both security and advertising use cases. It restarted classic customer clusters to patch CVE-2026-114938, a remote-code-execution flaw affecting specified OpenJDK 17.0.6-and-later systems with AVX-512 VBMI; serverless runtimes receive the fix automatically. The company reports no evidence of unauthorized access and expects no meaningful performance impact. Its CustomerLake Conversions product lets retailers send ad platforms purchase events selected using customer lifetime-value models, then compare campaign spending with customer value. Unity Catalog governs profiles and transfers, which leave audit trails; Epsilon can add identifiers through its identity graph.

Infrastructure and data access remain practical concerns. Together AI says its cache reaches warm reads of up to 3.4 GB/s per node on H100 systems—about 20 times faster than cross-cloud pulls—while cold reads are within 5% of direct downloads. Small-file warm-read latency falls from 320 and 680 milliseconds to 5 and 8 milliseconds, but gains depend on repeated reads and sufficient NVMe storage. Its T4 S3-compatible store is in limited preview. Separately, university libraries face high-volume remote scraping: one recorded more than two million bot queries in four weeks, with logs naming firms including Apple but not identifying most operators. The article recommends negotiating traffic and payment before crawling. The University of Nebraska–Lincoln will lead cybersecurity for the Fabric research project, focusing on protecting the integrity of its data and AI models.

Anthropic’s November 19, 2026 Nasdaq debut features across three reports. The company priced shares at $83, above the $70–$78 marketed range, raising $81.3 billion. Before trading, Nasdaq-100 tracker QQQ rose about 2%; indications at 11:38 a.m. ET pointed to an opening at $110.50, 33% above the IPO price, and a potential $2 trillion valuation. Analysts set targets ranging from $125 to $155, with Melius and Citizens expecting revenue above $400 billion by 2029. Anthropic has nearly $900 billion in computing commitments, including from Amazon and Google. One analyst expects a modest relief rally and names Nvidia, Google and Broadcom as potential beneficiaries, but not a $2 trillion first-day market value. Despite the tech rebound, the VIX stood at 25.46, and the Dow and small-cap shares rose less; investors were also watching a Federal Reserve announcement and Friday consumer-sentiment data. A separate opinion piece says AI’s electricity demand is drawing Big Tech toward traditional power companies, but its supplied discussion of Amazon contains no further specifics.

Debates over AI’s pace and governance continue. At Y Combinator, Garry Tan reportedly spends $100,000–$200,000 a year on AI tokens. In a discussion about slowing development, Hanson argued that intelligence is multidimensional, proposed liability insurance as a test, and warned that global control of computing could be captured. Jacob argued that improving performance on verifiable tasks signals progress and that $1 billion in liability coverage cannot address risks worth $10 trillion; he predicts a hundredfold speedup by late 2027 could automate AI research. Hanson said governments could restrict military knowledge without regulating all machines. In Washington, Rep. Don Beyer outlined a proposal requiring developers of the most powerful systems to use safety frameworks, report incidents and face external reviews; Sen. Mark Warner separately introduced a bill requiring public companies to disclose more about AI financing and oversight.

Other reports highlight organizational risks and oversight. A fake job interview helped attackers compromise Binance employees before an October cryptocurrency theft valued at $6.3 billion; the broader attack involved compromised devices, stolen credentials and misuse of an internal AI tool. The article recommends controlled coding tests and stronger coordination among recruiting, IT and security teams. KPMG’s audit-agent trial used 100 agents to review a major company’s financial statements; its UK audit head said they asked “better questions” than experts, but staff had to intervene, and the technology cannot yet handle thousands of audits. KPMG plans to expand the project to 20 more companies, while the UK Financial Reporting Council cautions against weakening auditor independence. OneTrust’s Fall ’26 release adds AI-system and risk maps, dashboards for usage and token consumption, and Deep Research AI for third-party assessments; the company says this can cut completion from weeks to minutes. It also adds protections against automated abuse of privacy request forms.

Finally, The Information reports that Microsoft executives told OpenAI they found no evidence of AI model distillation, while ByteDance is discussing about $500 million in additional cloud spending—developments described as straining Microsoft–OpenAI ties.

Key Takeaways

  • Microsoft expands Copilot with Work IQ, Auto model selection, agent identity controls and new work-measurement tools; Autopilot is rolling out to Frontier customers.
  • Linq’s a16z-led Series B backs messaging infrastructure for AI agents across Apple Messages for Business, RCS and WhatsApp; its API is used by more than 1,000 teams.
  • OpenAI demonstrated Codex Security reporting only P0 and P1 issues inside GitHub pull requests; pricing and rollout timing remain undisclosed.
  • Lovable says Rust-based development-server changes cut preview startup from 40 seconds to three, while checkpoint restoration takes under a second.
  • Databricks patched CVE-2026-114938 by restarting classic clusters; serverless customers receive patched runtimes automatically. Its CustomerLake Conversions targets ad events using customer lifetime value.
  • Together AI reports warm cache reads of up to 3.4 GB/s per H100 node, but benefits depend on repeat reads and NVMe capacity; university library scraping has reached more than two million queries in four weeks at one library.
  • Anthropic priced its IPO at $83 and raised $81.3 billion. Early indications put shares at $110.50, while analyst targets ranged from $125 to $155; market concerns remained elevated.
  • Anthropic has nearly $900 billion in computing commitments, including partners Amazon and Google. A separate report says ByteDance is discussing about $500 million in additional cloud spending as Microsoft–OpenAI ties face strain.
  • AI security and governance concerns span a $6.3 billion Binance theft linked to a fake interview, KPMG’s human-assisted audit-agent trial, and proposed US rules on safety practices and corporate AI disclosures.
  • The University of Nebraska–Lincoln will lead cybersecurity for the Fabric research project, while OneTrust’s Fall ’26 release adds AI-risk mapping, monitoring dashboards and automated third-party research.

Microsoft builds Copilot into the full workday

Microsoft is expanding Copilot with a Home page where employees can use Chat and Cowork, and an Auto feature that chooses models and effort for each task. Its Work IQ system finds useful work information across Microsoft 365, which holds more than 19.5 exabytes of data, including 90 billion emails and 15 billion Teams messages. Copilot can help create Word, Excel and PowerPoint files, while Autopilot agents get their own Microsoft Entra Agent ID and manager. Autopilot is rolling out to Frontier customers, Code is still weeks from general availability, and the Today command center is in private preview. Microsoft also reported new Dynamics 365 skills, use of Cowork by more than 5,000 Lilly employees, and an Agent 365 Business Analytics tool that measures work completed.

Linq raises Series B to put AI agents in messaging apps

Linq raised a Series B led by a16z, with Kimberly Tan, Josh Elman and Kenan Saleh backing its messaging technology for AI agents. The company connects businesses to chat apps through messaging tools that handle delivery, identity and scale. Linq is approved to use Apple’s Messages for Business and also supports RCS, WhatsApp and other channels. DoorDash shoppers can find food, build a cart and check out in a single message thread. Linq says agents make up more than half of internet traffic and that its API is trusted by over 1,000 teams; the company was founded by Elliott Potter, Patrick and Jared, and is hiring.

OpenAI demo brings Codex Security scans into GitHub pull requests

OpenAI demonstrated Codex Security reviewing scan results inside a GitHub pull request. Previously, Codex Security focused on the openai/codex-security command line tool and TypeScript SDK, which can find, investigate and fix code vulnerabilities. In GitHub, Codex flags only P0 and P1 issues, according to ChatGPT Learn docs, keeping comments focused on high-priority risks. The demo page did not list pricing, limits or a rollout date.

Heavy AI scraping is knocking university library sites offline

University library systems are struggling with remote, high-volume AI scraping that uses complex queries without requiring a login. Lewis said one library recorded more than two million bot queries in four weeks, linked to firms including Apple. The traffic logs identify brand names, but do not show who operates most of the bots. In October, Lewis said CP24 reported outages at university library websites. The article argues that organizations seeking library data should negotiate traffic levels and payment before crawling.

Lovable removes its 30-minute limit for AI agents

Lovable co-founder and CTO Fabian Hedin described how the company supports longer-running agents by restoring computer checkpoints in less than a second. A starting checkpoint includes the operating system, system services and a ready development server, and takes about 230 milliseconds to copy. Lovable also rewrote its development server in Rust, cutting preview startup from 40 seconds to three seconds while making code start twice as fast in 6.5 times less memory. Agents check work during each turn for build, lint, type and runtime errors, then use production behavior, A/B tests and business measures between turns. Hedin said the agents can access browsers and GitHub, connect to tools for hosting and other services, and currently process over a billion tokens per minute; proactive business tools are planned for 2027.

Databricks restarts clusters to block Java4Shell flaw

Databricks restarted customer clusters to apply a fix for CVE-2026-114938 before the flaw’s public disclosure could lead to scanning. Red Hat says the flaw can let attackers run code remotely without logging in on OpenJDK 17.0.6 and later, on x86-64 systems with AVX-512 VBMI. Databricks expects the fix to have no meaningful performance impact and says it has seen no evidence of unauthorized access. Serverless customers receive patched Runtimes automatically, while classic cluster users must restart to get the update. Databricks is also contacting users with custom images.

Together AI cache speeds up warm reads, with limits

Together AI says its cache can deliver warm reads of up to 3.4 GB/s per node on an H100 system, about 20 times faster than cross-cloud pulls. The speed boost applies after files have been fetched once and stored on cluster NVMe; cold reads were within 5% of a direct download. Warm reads cut latency for small files from 320 and 680 milliseconds to 5 and 8 milliseconds. The article notes that the benefits depend on jobs rereading the same data and on enough NVMe storage, while AWS egress can cost up to $0.09 per GB. Together AI’s T4 S3-compatible store remains in limited preview, and ASTRA’s engineering deep dive has not yet appeared.

Databricks lets advertisers target customers by predicted value

Databricks introduced CustomerLake Conversions to help retailers send higher-value purchase events to advertising platforms. Teams can use lifetime value models already stored in Databricks to select conversions, then send them through Reverse ETL and Conversions API connectors. Campaign results return through OpenSharing integrations and Lakeflow Connect so teams can compare ad spending with customer value. Databricks says Unity Catalog governs customer profiles, brands control which data they share, and each data transfer leaves an audit trail. Epsilon can add identifiers to events through its identity graph, while Databricks presents the tool as a step toward its planned Infinity Campaigns approach.

Y Combinator leaders clash over slowing AI development

Y Combinator chief Garry Tan spends $100,000 to $200,000 a year on AI tokens, according to XRoute. In a debate about whether society should slow AI, Hanson argued that intelligence is not one-dimensional and said people mostly adapt to problems. He proposed liability insurance as a test and warned that global control of computing could be captured. Jacob argued that growing model abilities predict progress on tasks that can be checked, and warned that $1 billion in liability coverage cannot address risks worth $10 trillion. Jacob expects a hundredfold speedup by late 2027 that could automate AI research, while Hanson said governments could restrict military knowledge without regulating all machines.

Anthropic IPO lifts AI stocks as market fears remain high

Anthropic began trading on the Nasdaq on November 19, 2026, after pricing its IPO above the marketed range. The debut helped lift the Nasdaq-100 tracker QQQ about 2% before the market opened, following a 3% drop in Nvidia shares the previous day. The rally was concentrated in technology stocks, while the Dow and small-cap shares rose less; the VIX stood at 25.46, signaling continued market concern. A strong debut could ease worries behind the Nvidia selloff, while a weak one could deepen them. Investors were also watching a Federal Reserve announcement scheduled for 2:00 p.m. ET and consumer sentiment data due Friday at 10:00 a.m. ET.

Anthropic IPO lifts tech stocks as market fear remains high

Anthropic began trading on the Nasdaq on November 19 after setting its IPO price above the range offered to investors. The debut helped revive AI stocks before the market opened, and a Nasdaq 100 tracker rose about 2%. The rebound came as the market’s fear gauge remained in a danger zone. The article says the recovery could still be undone by two key details, though the provided text does not identify them.

Analyst expects Anthropic IPO to lift AI stocks modestly

Paul Meeks of Freedom Capital Markets said Anthropic’s IPO could bring a relief rally for AI stocks after weeks of selling. Anthropic priced its shares at $83, above the marketed range of $70 to $78, and the offering raised $81.3 billion. Meeks said Nvidia, Google and Broadcom could benefit from their ties to Anthropic, while SpaceX may gain from its AI infrastructure spending. He called the IPO worth participating in based on expected revenue growth, but said he did not expect Anthropic to reach a $2 trillion market value on its first day of trading.

Analysts give Anthropic bullish ratings before Nasdaq debut

Brokerages including Needham and Melius Research began coverage of Anthropic with bullish ratings ahead of its November 19 Nasdaq debut. At 11:38 a.m. ET, indications showed the shares set to open at $110.50, 33% above the $83 IPO price, with the company on course for a $2 trillion valuation. Melius set a $155 target, Needham set $150, Citizens set $125 and Wolfe Research set $135. Melius and Citizens expect Anthropic’s revenue to exceed $400 billion by 2029, and the company has nearly $900 billion in computing commitments with partners including Amazon and Google.

AI power demand puts energy companies in the spotlight

A November 18, 2026 opinion article examines which companies may benefit as artificial intelligence drives demand for electricity. AI requires chips, data centers and large amounts of power, leading Big Tech firms to make deals with traditional power companies. The article urges investors to look beyond the headlines when considering who will profit. The provided text begins discussing Amazon.com but does not include further details.

University of Nebraska Lincoln to lead cybersecurity for AI research project

The University of Nebraska Lincoln will lead cybersecurity work for the Fabric for data-intensive and AI-driven science project. The project brings together universities and research institutions to help researchers find data, use AI services and conduct large-scale research. Its infrastructure aims to protect large datasets and AI models while supporting secure and reliable research. UNL will focus on the security and integrity of the data and models used in the project.

AI recruitment scams reveal new risks for company security

A fake job interview helped attackers compromise Binance employees before a $6.3 billion cryptocurrency theft in October. The wider attack involved compromised devices, stolen credentials and misuse of an internal AI tool, eventually giving attackers access to cryptocurrency wallet systems. The article warns that job interviews and coding tests can expose people to malicious software, while AI-generated images and voices can help fake applicants hide their identities. It recommends controlled environments for technical tests and stronger checks across recruitment, IT and security teams.

ByteDance AI sales put pressure on Microsoft and OpenAI ties

The Information reported that Microsoft executives told OpenAI they had found no evidence of AI model distillation. The report also says ByteDance is discussing about $500 million in additional cloud spending. These developments are placing strain on ties between Microsoft and OpenAI.

KPMG trial finds AI auditors ask strong questions but need human help

KPMG tested AI agents in Project Exponential to see how they compare with human accountants. One hundred agents reviewed financial statements for a major company, and UK audit head Cath Burnet told the Financial Times that they asked “better questions” than KPMG experts. The trial had mixed results, with staff stepping in when the AI struggled. Burnet said the technology cannot yet handle thousands of audits and many companies lack clean data, but KPMG plans to expand the project to 20 more companies. The UK Financial Reporting Council warned that AI can improve audits but must not weaken auditor independence.

Beyer proposes AI safety rules as Warner targets company disclosures

Rep. Don Beyer and Sen. Mark Warner introduced separate AI measures in Washington on Thursday. Beyer released an outline for bipartisan legislation that would require developers of the most powerful AI systems to use safety frameworks, report incidents and face outside reviews. The outline is only a blueprint, and lawmakers have not written its legislative text. Warner introduced a bill that would require publicly traded companies to disclose more about how they finance and oversee AI.

OneTrust Fall 26 release adds AI tracking and risk management tools

OneTrust’s Fall ’26 Release adds an interactive view of links among AI systems, models, agents, projects, risks, controls and workflows. New dashboards summarize AI use, risks, evaluation coverage, discovered outbound personal data and token use. The update also uses Deep Research AI to gather public information for third-party risk assessments, which OneTrust says could cut completion time from weeks to minutes. Privacy teams gain protections against automated abuse of data request forms, while updates to consent management and risk reporting bring related tasks together and help teams understand and prioritize risks. OneTrust is sharing availability details through an overview webinar.

Sources

NOTE:

This news brief was generated using AI technology (including, but not limited to, Google Gemini API, Llama, Grok, and Mistral) from aggregated news articles, with minimal to no human editing/review. It is provided for informational purposes only and may contain inaccuracies or biases. This is not financial, investment, or professional advice. If you have any questions or concerns, please verify all information with the linked original articles in the Sources section below.

Microsoft Copilot AI agents Microsoft 365 AI model selection AI messaging infrastructure OpenAI Codex Security software development Databricks AI advertising AI infrastructure data caching web scraping Anthropic IPO AI investment AI governance AI safety cybersecurity AI auditing AI risk management Microsoft–OpenAI relationship

Comments

Loading...