BriefTechNews
Optimizely Breach Turned 178 Websites Into Malware Dispensers

Optimizely Breach Turned 178 Websites Into Malware Dispensers

7 min read · 10 sources

TL;DR
  • Optimizely breach injected malware into websites of at least 178 organizations, targeting crypto wallets and payment data
  • Microsoft's Windows endpoint security platform and administrator protection are now generally available at Ignite 2026
  • Copilotforce integrates Salesforce with Microsoft Copilot, Teams, and Fabric with 38+ prebuilt sales skills
  • Snowflake's Open Data Sharing lets external consumers query Iceberg tables without a Snowflake account
  • Microsoft's API specs cover 90 of 178 services, with a Kin Score spread of 7.1 matching Amazon's 7.5

The biggest story today isn’t an AI announcement - it’s a supply-chain attack that turned the web’s most-trafficked pages into malware dispensers. Attackers exploited Optimizely’s experimentation platform to inject malicious JavaScript into websites belonging to at least 178 organizations. The payloads were hunting cryptocurrency wallets and payment details, and Optimizely says it has patched the vulnerability, removed the injected code, and revoked tokens.

If you run any site with a third-party experimentation or analytics tag, this is the nightmare scenario: a trusted vendor’s script becomes the attack vector, and you have zero visibility until the damage is done. It’s also a reminder that your security posture is only as strong as the least-trusted script you load on every page.

Meanwhile, Microsoft used Ignite 2026 to ship a wave of Windows security and AI-agent controls, and Microsoft and Salesforce announced a joint offering called Copilotforce. Here’s everything that matters.

Attackers used Optimizely’s experimentation software to push malicious JavaScript through websites belonging to at least 178 organizations.

Windows Gets Tighter Leashes on AI Agents

Source: blogs.windows.com ↗

Microsoft’s Ignite 2026 announcements are aimed squarely at the agent era. The Windows endpoint security platform and Windows administrator protection are now generally available, with Windows baseline protection in preview. Microsoft Execution Containers (MXC) gained new containment capabilities, and there are experimental hardware-isolated micro VMs.

The headline for IT teams: generally available learning modes let you observe agent behavior before enforcing restrictions. That’s a meaningful shift - instead of blocking agents cold or letting them run wild, you can watch what they actually do in production, then tighten the screws. Intune is also previewing automatic updates for applications employees install themselves, which closes a long-standing patch gap. And GitHub Copilot Fusion can now route coding work to local models on developer PCs, which matters if you’re trying to keep source code off external APIs.

The operational upside is real: Microsoft claims Autopatch now delivers updates 40% faster. For anyone who has spent a Tuesday night babysitting a fleet-wide update, that’s the number that pays for the whole keynote.

Copilotforce: Salesforce Data, Microsoft Front End

Source: news.microsoft.comource ↗

Microsoft and Salesforce announced Copilotforce, an integration that connects Salesforce with Copilot, Teams, and Fabric. The centerpiece is “Salesforce in Copilot,” a plugin in public preview with 38+ prebuilt sales skills. Employees can retrieve customer information and take supported actions under their existing Salesforce permissions - no more context-switching between a CRM and an AI assistant.

The infrastructure detail worth noting: two-way Zero Copy integrations between Salesforce Data 360 and Microsoft Fabric. That means you can query data across both platforms without duplicating it, which kills the classic “which copy of the customer record is canonical” problem. Accenture is the design and implementation partner and the pilot customer, which tells you where this is headed first.

For engineers, the appeal is that permissions carry over. The plugin respects existing Salesforce security models, so you’re not building a new authorization layer to bolt AI onto your CRM data.

When AI-Generated Data Becomes Tomorrow's Ground Truth

Source: oreilly.com ↗

An O’Reilly analysis lays out a systemic risk that should worry anyone running AI-integrated workflows: AI-generated conclusions can lose their original context and become treated as established facts. The example is a supplier-management system where an AI agent misreads a quality certificate’s scope. Even after a human corrects the record, downstream dashboards and reports keep acting on the old, incorrect finding.

The core problem is that correcting a value doesn’t correct its influence across caches, reports, and workflow state. Agents make this worse by generating conclusions that later get treated as independent evidence. The proposed fix is “decision lineage” - linking AI-generated assertions to their source evidence, governing workflows, and downstream decisions.

For engineers, this is the data lineage problem you already know, but with a new twist: the bad data isn’t just a wrong number in a database, it’s a wrong conclusion that an agent will cite as ground truth for the next decision. If you’re building agentic systems, you need to design for correction propagation, not just data entry.

Teams Gets Impersonation Protection

Source: office365itpros.com ↗

Microsoft is adding three new capabilities to Teams to fight social engineering in meetings. “Impersonation Protection” rolls out mid-December 2026 and will detect and flag potentially deceptive participants and organizers. The “Facilitator” agent gains a new protection skill in late January 2027 that scans meeting transcripts for threats, and there’s a Microsoft Defender data table for forensic investigations.

The caveat that should grab your attention: admins have no visibility into the signals used for detection and may not be able to disable it. That’s a classic trade-off between security and control. If the detection logic produces false positives - flagging a legitimate vendor as a potential impersonator - you have no way to tune or override it. Worth planning for before you roll it out to your sales team.

Snowflake Opens the Door to Non-Snowflake Consumers

Source: snowflake.com ↗

Snowflake launched Open Data Sharing, which lets data providers share Apache Iceberg tables with external consumers using Iceberg REST Catalog-compatible clients. No Snowflake account required, no scheduled data exports. This is a direct answer to the “we’d share this data, but our partners don’t run Snowflake” problem.

For teams living in mixed technology environments, this kills the export-and-email pattern. You publish once to an Iceberg REST Catalog, and any client that speaks the protocol can query live data. It complements existing Snowflake sharing pathways rather than replacing them, which means you can keep your internal Snowflake-native sharing and open up specific datasets to the outside world.

The API Source-of-Truth Scorecard

Source: apievangelist.com ↗

An API Evangelist analysis compared the API estates of Google, Amazon, and Microsoft, and the numbers are stark. Amazon’s Smithy models cover 186 of 199 API members, Microsoft’s specs cover 90 of 178, and Google’s Discovery documents cover 126 of 161. Microsoft has the most generated operations at 49,891.

The finding that matters: consistency (measured by Kin Score spread) tracks with the completeness of the machine-readable source, not the company. Microsoft’s spec-built members have a spread of 7.1, matching Amazon’s 7.5. The lesson for anyone building or consuming APIs: a complete, machine-readable source of truth directly correlates with consistency and agent readiness. If your API docs are hand-maintained and drift from the implementation, your agents will feel it.

Open-Source E-Reader Project vs. the AI Flood

Source: theregister.com ↗

Crosspoint Reader has passed 300,000 installations, bringing open-source firmware to tiny, inexpensive e-readers. The growth is a win for the project, but it also illustrates the pressure AI-generated contributions put on volunteer maintainers, who must review submissions and manage feature requests within tight hardware limits.

This is the quiet story of 2026: LLMs let anyone generate a pull request, but they don’t generate the context to review it. Maintainers are drowning in plausible-looking code that needs the same scrutiny as human-written code, with none of the human’s ability to explain why they wrote it that way. If you maintain an open-source project, you’ve felt this. If you’re generating contributions with AI, consider whether you’re helping or adding to the review queue.

SAP Swallows TechWolf for Workforce Grounding

Source: news.sap.com ↗

SAP completed its acquisition of TechWolf, an AI work intelligence platform. TechWolf’s proprietary context graph becomes SAP’s grounding layer for agent queries related to work, skills planning, and talent management within SAP SuccessFactors. The pitch is that grounding agents in a context graph makes token usage more efficient, lowering the cost of deploying workforce agents.

For engineers running HR automation, the interesting bit is that a context graph is the difference between an agent that hallucinates a skills assessment and one that pulls from a structured, maintained view of the workforce. Whether SAP actually delivers on that is another question, but the architecture is the right one.

RHEL GitHub Runners Go GA

Source: developers.redhat.com ↗

Red Hat announced that official GitHub-hosted runner images for RHEL 9 and RHEL 10 are now generally available. Teams using GitHub Enterprise Cloud and GitHub Enterprise Server can run CI/CD pipelines natively on RHEL, set up via “New larger runner” in organization settings.

The value is eliminating the “works on my Ubuntu container, breaks in prod on RHEL” class of bugs. Library versions, compiler behavior, and SELinux policies differ between distros, and testing on the wrong OS hides real issues. Clean base OS images on 64-bit Linux mean your tests run on the same OS as your deployment targets. It requires a GitHub organization account, but for RHEL shops, this closes a real gap in the CI/CD pipeline.

Get the brief

Liked this one? Pick the beats you care about and get just those, five minutes a morning. Free, no hype, one click to leave.

About Me Author

My name is

BriefTechNews

A daily digest of what actually moved in AI, tech, crypto and fintech, assembled and written with AI, with every item linked to its source. Read More
Tags

You May Also Like