The Agentic Enterprise. AK · Wednesday, November 18, 2026 · 8 min read | Wednesday, November 18, 2026 The attacks came in through the suppliersExpedia, Infoblox, Uber and Visa have all traced recent incidents to a supplier, and on Tuesday two poisoned releases of Microsoft's agent SDK briefly sat on PyPI. Governments are on the list too. Japan is spending $63 million on an emergency response after dozens of companies and agencies were breached, and Belgian prosecutors are investigating attacks that sent supermarket refrigeration in several European countries running hot. In the four corporate cases, attackers came in through credentials or data that a vendor already held, and three of those vendors still have not been publicly named. | The Big Story Security Most of the doors were someone else's| E | xpedia said on November 12 that attackers reached its systems through a compromised third-party data platform. Researchers at Wiz and Sentra documented an October attack through Qubole, a onetime Expedia vendor whose former customers never revoked its access, according to Sentra; neither Expedia nor the researchers have tied Expedia's breach to it. Infoblox, which sells DNS and identity services, said attackers reached company information through a data migration vendor; two people told TechCrunch the vendor handles SharePoint migrations. Uber said compromised credentials at its inference provider, Fireworks AI, exposed license data for at least 7,870 drivers in three states. Visa said an incident claimed by the n0n extortion group began at a provider in Asia Pacific and didn't touch its own systems. Skift · TechCrunch · Cybernews · Cybernews |
On Tuesday, two releases of agent-framework, Microsoft's agent SDK on PyPI, carried a credential stealer for about 45 minutes. It ran at every Python start and went after Azure, pipeline and GitHub tokens. The package gets about 190,000 downloads a month. Aikido · The Hacker News Public bodies are taking hits as well. Japan announced a ¥10 billion ($63 million) emergency package on November 6, and JPCERT/CC confirmed signs of AI agents in the attacks. On November 2, refrigeration at more than 100 European shops ran hot in what the EU's agriculture commissioner called a "deliberate attempt to disrupt" food supply. Reuters reported that the Pentagon treats August freezer failures at two US military sites as a suspected hack. Insurance Business · Al Jazeera In each corporate case where the company has named a cause, the way in was a supplier's access or data. How the agent-framework attacker got publishing rights is still unknown. Every vendor on the list holds a key. The Spearhead Take List every vendor, former vendor and package that holds standing access to your data or build systems, and revoke what nobody uses. Swap long-lived keys for short-lived credentials. Pin agent SDKs by hash behind an internal mirror. If any environment pulled agent-framework 1.26.1 or 1.26.2, rebuild it and rotate every secret it could reach. |
Sources: Skift · TechCrunch · Cybernews · Aikido · Al Jazeera | The Obvious & The OverlookedWhat is easy to see, and what is easy to miss. The Obvious Anthropic prices tonight. The range is $70 to $78 a share, orders reached about six times the shares on offer, and trading is expected Thursday. Source: StockAnalysis. | A robot maker raised a half-billion-dollar Series A. Genesis AI raised a $500 million Series A at a $3.5 billion valuation for Eno, a wheeled two-armed robot. Source: The Next Web. | Investors are backing agents for bankers. Model ML raised a $250 million Series B at a $1.5 billion valuation, co-led by General Atlantic and QED, for agents that work inside bankers' existing systems. Source: Tech Funding News. |
|
The Overlooked Data centers are already on the grid's books. A Harvard-led study puts 403 US hyperscale sites at 68 to 99 TWh a year, about 1.8% of US electricity, with roughly 54% from fossil fuels. Source: EurekAlert. | Inference is moving into old buildings. With power and cooling gear on 18- to 24-month lead times, SambaNova is pitching air-cooled racks for telco and colocation sites that can't take 120 kW GPU racks. Source: SiliconANGLE. | Token bills are driving model choice. Brad Maltz of Dell said, "We didn't expect to spend $10,000, $20,000 per person per month," on frontier tokens, in a Dell-sponsored interview. Source: SiliconANGLE. |
|
|
| Moving PiecesOpenAI cuts a price and raises a risk rating, Nvidia beats without a permanent CEO, Microsoft ships its own models, and Ignite and Cerebras change where agents run. Product OpenAI makes Astra cheaper and riskier on paperGPT-6.2 Astra costs $8 per million input tokens and $40 output, 20% below its predecessor; Sol holds at $2 and $10. Astra is the first externally released OpenAI model rated "High" for AI self-improvement, and both models sit at the "Critical" cybersecurity threshold, so OpenAI tightened default cyber help. Enterprise and Edu workspaces need an admin to switch them on. Early reviews are mixed. Before routing work to it, find out who in your company flips that switch. Sources: The Decoder · Unite.AI · Latent Space | Compute Nvidia beats with a CEO search still openRevenue reached $115.8 billion, up 103%, with data center at $107.6 billion. Nvidia guided to $130 billion next quarter, assuming no data center compute sales to China. Colette Kress is interim CEO and CFO; the board is running an investigation through Wachtell Lipton, and a source told CNBC it is not weighing Jensen Huang's return. The board also added $150 billion to its buyback. Vera Rubin racks are running at five cloud partners. Nobody yet knows who will set Nvidia's product and pricing strategy once the search ends. Sources: NVIDIA · The Next Web | Platforms Ignite's agent controls trail its agentsOffice Watch counted 215 features in Microsoft's Ignite news. Seventy-six are generally available; about half are previews or Frontier-only. Autopilot, an always-on agent with its own Entra identity, is Frontier-only. Agent 365 Cost Management, with spending limits and alerts, is in preview, and an action center that finds ownerless agents reaches private preview in December. Budget for the agents on the general-availability list. Plan your governance around the controls that are actually shipping. Sources: Office Watch · Microsoft · SiliconANGLE | Infrastructure Cerebras hands half the job to AMDCerebras now splits each request: AMD Instinct GPUs process the prompt, and its CS-3 systems generate the output. On a shared GPT-OSS-120B endpoint, it says the setup produced five times the tokens per megawatt of CS-3 alone, with no loss in speed. AWS and Fireworks AI are early partners, and Cerebras plans hundreds of megawatts by the end of 2027. Prompt processing and output generation now run on different chips. Ask providers which chip runs each half. Sources: Cerebras | Models Microsoft prices its coding model under the labs it resellsMicrosoft launched MAI Code 2 and MAI Code 2 Flash in GitHub Copilot. On Microsoft's own four-benchmark test, Code 2 averaged 63.2%, against 64.1% for GPT-6.1 Sol and 65.3% for Claude Opus 5.5, at $0.31 a task versus $0.76 and $6.31. Those costs use promotional prices that double after May 17, 2027. MAI Cyber 2 Flash moves into Microsoft's security tools in mid-December. Microsoft also released MAI Image 2.7 and the Decision-1.1 router. Run your own code through it before the promotion ends. Sources: InfoWorld · TestingCatalog |
| On the RadarEight signals worth tracking, one line each. Security Microsoft made Defender runtime protection for local AI agents generally available, but it requires Defender for Endpoint P2. Office Watch |
Operations Genpact made Onyx available, an agent platform where agents recommend actions for approval and earn autonomy as they prove reliable, starting with accounts payable. Constellation Research |
Data LanceDB raised a $57.5 million Series B co-led by Bessemer and Innovation Endeavors for a multimodal data lakehouse built on its open Lance format. SiliconANGLE |
Outages Apple's iCloud sign-in, App Store and Apple Pay went down Tuesday; Apple had not given a cause when 9to5Mac reported it. 9to5Mac |
Compute Azure Cobalt 200 VMs are generally available, with Microsoft claiming up to 50% better CPU performance than Cobalt 100 and 24 regions by year-end. Microsoft |
Infrastructure Dell launched OpenManage Rack Scale, which runs AI clusters at rack level and supports 100,000 devices in its first release. SiliconANGLE |
Data Wherobots put Waldo in public preview, an agent that builds spatial data pipelines from plain-language prompts, at no extra charge on its Professional Edition. SiliconANGLE |
Agents Cognition put Devin Bots in early access, managers that split long coding projects across Devin agents, each with its own budget and scoped repository access. Fortune |
|
| Quick HitsThe wider field, one line each. Funding figures and product claims are the companies' own. Agent tooling | 01 | Microsoft IQ is generally available, with no new SKU required. Microsoft | | 02 | Microsoft's Agent Optimizer is generally available; it proposes agent configuration changes and tests them against evaluations. Microsoft | | 03 | Copilot Code, which builds apps from plain language, is promised for general availability in December. Office Watch | | 04 | Viva Insights agent reports require at least 50 assigned Copilot licenses. Office Watch |
|
Security and patches | 05 | Oracle shipped Java patches Tuesday; early reports describe an unauthenticated remote-code flaw in Java 17, which Aikido rates 9.2 under CVSS 4.0, and Aikido advises treating Java 21, 25 and 27 as in scope. Aikido | | 06 | Dynamic Agent Risk, which scores agents by runtime signals and blast radius, entered preview with Check Point, Darktrace and Zenity feeding it. Microsoft |
|
Money and research | 07 | Tesorai raised a $5.3 million seed from Perplexity Fund, IQT and South Central Ventures to mine raw lab-instrument data for drug discovery. SiliconANGLE | | 08 | Promise Robotics raised a C$40 million Series B led by White Star Capital for robots that build prefabricated homes. BetaKit |
|
| From Spearhead · AI Bootcamp Get your leadership team AI-ready, in a day.The AI Bootcamp is a one-day, in-person session designed for your industry, your organization, and your use cases. Your leaders understand AI, build AI agents, envision outcomes, and start building a new work culture. |
| The Number 64.5% The share of list-price spend that went to agent runs that failed or timed out, across 27,280 coding-agent runs in Runta's FrontierHarness benchmark. The same study found the model explained 92.8% of the variance in cost per task and the harness only 2.3%. The model sets most of the cost, so measure failed runs before you tune the harness, and re-test the harness whenever you switch models. Runta | Counter-Signal Security The loudest outages weren't attacksThe week's scariest headlines were mostly self-inflicted. NSA's classified internal network went down Tuesday during what the agency called "scheduled routine maintenance." The FAA's four-hour-45-minute nationwide ground stop on November 13 likely traced to excess network data from a site where its contractor was installing a new router, and the FAA says it has "no indication that this was a cyberattack." JPMorgan Chase blamed its outage on a software bug. ABC News · FAA · Cybernews Director of National Intelligence Jay Clayton put it down to upgrades and patching, and Sen. Mark Warner said AI-driven pressure to patch many systems at once is straining systems that weren't built for that pace. That caution has a cost. Pinning packages and cutting vendor access also slow the patches you need, so set a fast path for urgent fixes. ABC News Sources: ABC News · FAA | From the Field The detail that stays with me this week is Sentra's finding that Qubole's former customers never revoked its access. When we review enterprise AI programs, we usually find access that outlived the contract. A pilot ends, the contract lapses, and the service account, the API key or the data-share keeps working because nobody owned the off-boarding step. AI work makes that list longer. Every inference provider, evaluation tool, data-labeling firm and agent SDK needs a credential or a copy of the data, and most arrive through experiments rather than procurement. Uber's driver data left through its inference provider. Tuesday's stealer arrived through a package a developer installs in seconds. The fix is the discipline most security teams already apply to production systems: a list of who holds access, with an owner and an expiry date for each entry. The work is extending it to inference providers, eval tools and SDKs that arrived through pilots. Let's create value with AI, AK | The Agentic Enterprise Know more about AI than 95% of your peers. Every morning. Anthropic, OpenAI, Google, Microsoft/Azure, AWS, NVIDIA, Databricks and Snowflake are Spearhead technology partners; this edition covers several of them, including a compromise of Microsoft’s Agent Framework package, Microsoft’s own models and Anthropic’s IPO, and applies the same scrutiny to each. This edition was produced using Anthropic’s Claude with editorial direction from AK. Sources are linked inline. © 2026 Spearhead. All rights reserved. |
|