In the past two weeks, vendors serving Chase and Mastercard were breached. An IT provider for Irish credit unions learned more than a month late that attackers had reached member data. Microsoft 365 email went down for most of a business day, CISA added AI tools to its list of actively exploited software, and federal agencies ran into multi-day outages of their own. Amid that 10-day wave of disruptions, change is now occurring right at the top of U.S. cybersecurity, with the White House moving to a new acting National Cyber Director on November 19. None of this needed exotic defenses. Each incident traces back to ordinary gaps in people, tools or IT operations, and those gaps can be found and fixed before someone else finds them.
Third-party vendor breaches, major outages and AI tool risks: the recap
Bank and card vendors breached. JPMorganChase said “the attackers accessed certain marketing data held by those vendors, which may include some customer names and email addresses,” and that its services weren’t disrupted. Mastercard said a provider exposed a limited number of cardholder email addresses. Visa said an incident at a third-party provider in Asia Pacific didn’t involve its systems or card numbers. Cybernews researchers reported that about 16,000 Chase and more than 6,000 Visa customer records were leaked (reported, not confirmed by the companies).
A patched hole that was already used. In Ireland, an IT provider serving 43 credit unions closed a vulnerability on October 6 during a routine assessment. Its first investigation found no breach. On November 15 it found attackers had reached saved payee details.
Microsoft 365, November 12 to 13. A lock-contention defect introduced in a July update for Microsoft 365 networking gateway servers, combined with extra Teams failover traffic and nearly 14% of gateway capacity already offline, cascaded until more than 80% of gateway servers stopped responding and Exchange Online was disrupted for just under 24 hours. In its preliminary post-incident report, Microsoft acknowledged that its lab testing missed the bug because the test environment didn’t match production hardware: “The contention issue is worsened on machines with more cores and newer hardware; whereas our load testing tended to focus on weaker hardware with fewer cores.” Microsoft called the outage “a technical issue not related to cyber or AI / agents,” and its final report was due November 20 (we’ll update this post when it’s published). Even setting AI aside, the report highlights a broader blind spot for customers: Microsoft does not publicly report what share of its code updates and patches are AI-written today (CEO Satya Nadella’s April 2025 estimate of 20% to 30% of repository code remains its last company-wide figure), nor does it disclose the details of its internal patch testing before updates ship. That’s why BTI explicitly tests and whitelists updates as part of our routine managed patching and Protection Level Agreements for patching and maintenance rather than pushing vendor updates straight into client production.
FAA ground stop, November 13. A newly installed router at a remote site was a suspected source of data that flooded the network during a midday change. The Transportation Secretary said the problem “was not identified quickly enough,” and on November 19 FAA Administrator Bryan Bedford said the outage wasn’t tied to modernization, pointing instead to an “archaic” federal telephone network that needs replacement. The FAA hasn’t announced a final root cause yet.
USDA multi-day outages after a security incident, November 16 to 19. Nextgov/FCW reported Thursday that the U.S. Department of Agriculture took numerous applications offline starting Monday after its Cybersecurity and Privacy Operations Center “uncovered a security incident.” According to an internal email from an acting assistant CIO, staff had to “reset all database passwords in all environments,” forcing application resets that disrupted farm loan, crop insurance and financial aid tools. A Thursday bulletin warned intermittent outages could continue through the week, and some staff also ran into VPN issues. Officials haven’t said who was behind it or whether data was taken.
National Security Agency internal network outage, November 17 to 19. The NSA’s classified internal network went down Tuesday during maintenance, prompting staggered arrivals and leaving some nonessential staff working from home for a third straight day on Thursday, according to NBC News. The NSA called it a “temporary degradation during scheduled routine maintenance” and said mission-critical capabilities stayed operational. Director of National Intelligence Jay Clayton confirmed it wasn’t a cyberattack and linked it directly to urgent patching of critical PHP vulnerabilities uncovered with AI tools last month, saying the patching “caused some of these systems to not come back up when you patch them as expected.” In a November 18 White House statement, Clayton and National Security Adviser Marco Rubio said there is no indication of a coordinated cyberattack across the recent tech outages, pointing instead to widespread urgent patching across interconnected legacy systems.
Change at the top of U.S. cybersecurity, November 19. Even at the top of federal cyber leadership, change is now underway. Following the 10-day string of banking, aviation, cloud and agency disruptions, Fox News confirmed Thursday that National Cyber Director Sean Cairncross is leaving his post and Pentagon Chief Information Officer Kirsten A. Davies is stepping in as acting National Cyber Director. A source close to the White House told Fox News the leadership change followed concerns about Cairncross’s cybersecurity credentials, while an administration official separately said he did not want to serve another two years. Davies called for modernizing older infrastructure and building resilience into core systems, and former White House CIO Theresa Payton underscored the operational takeaway for businesses rushing to patch: test updates, maintain backup systems and be ready to roll changes back.
AI tools on the exploited list. From November 9 to 18, CISA added 17 vulnerabilities to its Known Exploited Vulnerabilities catalog. They include an AI gateway (LiteLLM), GPU cluster software (GPUStack) and a common office storage system (Synology DSM). According to SANS, stolen credentials from the AI platform Fireworks AI led to the theft of nearly 150,000 Uber driver records, and Uber’s alerts never fired because the logins looked legitimate.
No one can promise zero breaches or zero downtime. No toolset stops a breach at someone else’s vendor or an outage inside Microsoft. What a toolset can do is make sure your own gaps are found first, so the impact on you is smaller.
This week’s patch list: the software hiding inside your systems
Some of this week’s most serious flaws are in software most businesses never chose to install. It ships inside servers, appliances and vendor products.
- nginx web server (“NGINX2Root”). On November 18, F5 disclosed five critical-rated nginx flaws that can let an attacker run code without logging in under certain setups, including CVE-2026-123007 when HTTP/2 is on. Researchers have published proof-of-concept code on GitHub. Its author says it works only under narrow lab conditions, but public code gives attackers a head start. Action: update nginx to 1.30.7 or 1.31.9 (NGINX Plus 37.0.8.1 or 37.3.0.1), or turn off HTTP/2 until you can. Ask your appliance vendors whether their products include nginx.
- VMware vCenter (CVE-2026-123890, rated 9.8). A login bypass, fixed November 19. Google’s threat researchers report that a suspected China-linked group used it for about two months, after first breaking into a remote support device. Action: apply Broadcom’s fix, limit who can reach vCenter and check for unexpected admin accounts. vCenter 7.0 has no public fix.
- Java (CVE-2026-114938, “Java4Shell”). Fixed November 17. It affects Java 17 and later on certain newer Intel and AMD processors. Attackers have moved past scanning: GreyNoise, watchTowr and Cloudflare now report probing and exploit attempts, though no confirmed break-ins so far. CISA urges organizations to patch immediately. Action: find every Java install, including Java bundled inside vendor products, containers and appliances, and update.
- LiteLLM AI gateway (CVE-2026-40217). CISA added it to the exploited list on November 18 and set a December 2 deadline for federal agencies. An attacker with admin access can run code through a guardrail testing feature. Action: update to LiteLLM 1.83.11 or later. Until you can, block the /guardrails/test_custom_code endpoint at your proxy or gateway and limit who holds admin keys.
- PHP. Updates on November 19 fixed seven critical flaws, including one an attacker can trigger without a login (CVE-2026-129496). Earlier AI-assisted PHP vulnerability discoveries already triggered the urgent patching wave linked to this week’s NSA internal network outage, a reminder to test PHP updates in a pilot group first. Action: update to 8.2.36, 8.3.37, 8.4.28 or 8.5.13, and plan to move off PHP 8.2 before its support ends December 31.
- Hosted email filtering. A hosted email filtering service used by many IT providers confirmed attackers reached one of its servers. It says mail kept flowing and it has no evidence customer data was accessed. Action: ask your provider whether your filter was affected.
- Website scripts. On November 16, an attacker inserted malicious code into some customers’ sites through a website testing tool. Action: keep a list of every third-party script on your website.
- Fake outage emails. Phishing posing as Microsoft 365 outage notices urged readers to “restore messages.” Action: check status only in the Microsoft 365 admin center.
Cameras, access control and intercoms run this software too. NVRs, video management servers, door controllers, portals and remote support tools often include nginx, PHP or Java, and each manufacturer decides when a fix ships. The vCenter case shows attackers moving from a remote support tool into the virtualization layer, where many video and security servers run. Keep management interfaces off the internet. Limit who can reach hypervisors and remote support tools, and ask each manufacturer whether your devices are affected. From networks, cloud and AI to cybersecurity, data governance and connectivity, BTI designs, sells, installs and commissions it all, including access control, alarm and camera systems that feed your security operations and business intelligence. That means one partner can check both your IT and your security systems.
New this week in AI security and governance
A trusted AI library turned into a credential thief. On November 17, two hijacked releases of Microsoft’s agent-framework library for Python (versions 1.26.1 and 1.26.2) were published to PyPI, the main public repository for Python code. They were live for less than an hour. According to Wiz and the Python security advisory, anything that installed them ran code built to steal Azure, CI pipeline and developer credentials and to request fresh cloud tokens. The advisory says any system that installed either version should be treated as compromised and every reachable credential rotated. Version 1.26.0 is safe. The lesson for every business: know which AI tools and code libraries your teams and developers install, and keep the credentials those tools can reach to a minimum.
An AI agent that reached too far. Reuters reports that during employee testing, Meta’s AI agent Muse mixed up conversations across accounts and surfaced data that wasn’t the user’s, including other people’s bank balances. One employee said it reminded her about a Signal message, though she said she never granted it access to other apps. Meta says employee testing exists to catch problems like these. It’s a clear example of why every AI tool needs limits on what it can see and do.
UK bank regulators publish AI cyber practices. On November 19, the Bank of England, the PRA and the FCA published effective practices for managing AI cyber risk at financial firms. They aren’t formal requirements, but they’re specific. They say AI now lets attackers exploit flaws “within hours or days rather than weeks or months.” They call for “execution speed” in fixing flaws, continuous monitoring of exposures, human oversight of AI, and “kill switches” so AI systems can be “paused, constrained or shut down safely.” The guidance also asks firms to review vendor contracts for AI risks and patching commitments. It fits the bank vendor story at the top of this post.
California proposes mandatory AI safety rules. On November 18, Governor Newsom released expert recommendations for mandatory AI safety and security standards, licensed independent auditors, and state power to order an emergency shutdown of a dangerous AI system. According to CalMatters, they would also shorten the window for reporting critical AI safety incidents from 15 days to 48 hours and limit using AI to build AI. These are recommendations, not law, and they mainly target large frontier AI developers. Legislators return December 7.
A new House AI Safety and Innovation Caucus. On November 19, 26 House members launched a bipartisan caucus, co-chaired by Reps. George Whitesides, Nathaniel Moran and Don Beyer. Its one-pager warns that AI-driven attacks pose “acute risks to critical infrastructure, sensitive personal data, and financial sector stability.”
What this likely means for your business. Expect more hearings, more headlines and more guidance on AI security in the months ahead. Even before new laws pass, published guidance tends to shape what counts as reasonable care. Frameworks like the NIST AI Risk Management Framework, state privacy rules and regulator guidance like the UK’s can become the yardstick a court or regulator uses to judge whether a business took reasonable steps. Ohio’s data protection law already gives businesses a legal defense if their security program follows a recognized framework such as NIST’s. And a November 2026 review of FTC cases by the law firm Loeb & Loeb found that vendor diligence and fixing vendor-flagged risks “are now expected components of a ‘reasonable’ privacy and security program.” After a breach, plaintiffs’ lawyers can point to that same guidance in privacy, PII and negligence claims and ask why a business didn’t follow it. This isn’t legal advice, so talk to your attorney about your own obligations. The practical takeaway is simple: document what you do to protect data and govern AI, and close the gaps that published guidance already names.
Privacy and compliance rules for small business: state laws, GLBA, HIPAA and more
Most businesses fall under more than one of these rules. In plain language, here’s who each applies to and what it expects. This is a summary, not legal advice.
State comprehensive privacy laws (23 to 24 states, depending on the tracker). According to the IAPP’s state privacy tracker, 23 states have passed comprehensive consumer privacy laws (some trackers, such as MultiState, count 24): California, Virginia, Colorado, Connecticut, Utah, Oregon, Texas, Montana, Delaware, Iowa, Nebraska, New Hampshire, New Jersey, Tennessee, Minnesota, Maryland, Indiana, Kentucky, Rhode Island, Louisiana, Oklahoma, Alabama and Vermont. Nineteen are already in effect. Louisiana and Oklahoma start January 1, 2027, Alabama on May 1, 2027 and Vermont on January 1, 2028. Who: businesses that handle personal data of enough residents of a state, with thresholds set by each law. What matters is where your customers live, not where your office is. What: honor rights to access, delete and opt out, limit data use and, in most states, assess risky processing.
State breach notification laws (all 50 states). Who: any business that holds residents’ personal information. What: notify affected people, and in many states regulators, after a breach. Deadlines and details vary by state.
GLBA and the FTC Safeguards Rule. Who: financial businesses that aren’t banks, such as mortgage lenders and brokers, finance companies, collection agencies, tax preparers and certain financial advisors. What: a written security program with a named Qualified Individual, risk assessments, access controls, encryption, MFA, change management and logging of user activity. If unencrypted customer information about 500 or more consumers is taken, you must notify the FTC as soon as possible and no later than 30 days after discovery.
FTC Act Section 5. Who: nearly every business. What: no unfair or deceptive practices, which includes keeping your privacy and security promises and not overstating what your AI does. The FTC’s 2026 cases have focused on companies that shared personal data in ways their privacy policies didn’t allow, including photos given to a facial recognition AI company, and on a company that didn’t fix security problems its own vendor had flagged.
SOX IT general controls. Who: public companies. What: management must assess its internal controls over financial reporting, and that includes the IT controls behind financial systems: who has access, how changes are approved and tested, and how systems are run and monitored.
SEC cyber disclosure. Who: public companies. What: report a material cybersecurity incident on Form 8-K within four business days of deciding it’s material, and describe cyber risk management and governance each year.
HIPAA Security Rule. Who: if you handle health data as a healthcare provider, health plan or a vendor to one. What: administrative, physical and technical safeguards for electronic health information, starting with a risk analysis. HHS has proposed an update, but it isn’t final, and the current rule still applies.
PCI DSS 4.0.1. Who: any business that takes card payments. It’s an industry standard, not a law. What: protect cardholder data. The 51 requirements that started as best practices became mandatory March 31, 2025.
California’s CCPA regulations on automated decisions. Who: businesses covered by California’s privacy law. What: rules that took effect January 1, 2026 add risk assessments, annual cybersecurity audits for some businesses and rights around automated decision-making technology. If you use those tools for significant decisions about people, you must comply by January 1, 2027.
Colorado’s automated decision law (SB 26-189). Who: businesses that build or use automated decision tools for consequential decisions such as hiring, lending, housing, insurance or healthcare. What: starting January 1, 2027, notice before use, an explanation after an adverse decision, a way to request human review, and three years of records.
What your business should do
People
1. Train staff to verify before they click. Leaked names and emails feed targeted phishing. Teach staff to treat unexpected calls, texts and emails about accounts or payments with suspicion.
Ask your provider for: short, regular phishing training and an easy way to report suspicious messages.
2. Name owners and map dependencies. Name an owner for each critical system and vendor. Keep a call tree that works without email or Teams.
Ask your provider for: a named contact, a clear after-hours escalation path and a vendor dependency map.
Tools
3. Build real redundancy. A second internet circuit with failover, backup power for network gear, a phone system that works when Microsoft 365 is down, and a separate Microsoft 365 backup.
Ask your provider for: a list of your single points of failure and alerts when a backup path goes down.
4. MFA, filtering and 24/7 monitoring. Enforce MFA for every user and admin. Use email filtering and link protection, and watch for logins that look legitimate but behave strangely.
Ask your provider for: written SLAs and around-the-clock monitoring with a team that responds.
5. Govern AI tools and vendor access. Keep a list of every AI tool and AI code library in use, patch AI software, and limit what each tool can reach. Ask each vendor what of your data it holds and who it shares it with. Share less, and retire keys and access when projects end.
Ask your provider for: an AI tool inventory and scheduled vendor and AI access reviews.
Processes
6. Patch fast, then verify. Patch everything on the list above, including software bundled inside appliances, cameras and access control systems. Don’t push vendor updates straight to every machine blindly: test and whitelist patches first, send them to a pilot group and schedule changes with a rollback plan. After patching an exposed system, check whether it was compromised before the patch.
Ask your provider for: a documented change process, patch pilot groups and post-patch compromise checks.
7. Write the plan and test it. Keep a BCDR plan with regular restore tests and a one-page “when email or internet is down” playbook.
Ask your provider for: scheduled restore tests with results you can see.
8 questions to ask your IT provider
Use these eight questions in your next IT review (and see our 5 questions to ask your IT provider on remote access tools for a deeper check on vendor access).
- Do our staff get regular phishing training?
- Who is our named contact, and what is the after-hours escalation path?
- What are our single points of failure?
- Is MFA enforced for every user and admin?
- Which vendors and AI tools hold or reach our data, and when was that last reviewed?
- Which of our devices, including cameras and door systems, run nginx, PHP or Java?
- After a critical patch, do you check for signs of compromise?
- When was our last restore test, and can we see the results?
One AI-enabled toolset, from AI-ready infrastructure to security
BTI delivers this as an AI-enabled a la carte, co-managed, or fully managed toolset customized for each client’s IT infrastructure and support capabilities. In one flow, it covers the six areas every business should manage proactively.
1. AI-ready secure infrastructure (AI, network management). Network capacity and redundancy, identity and MFA, organized and protected data, and governed access to AI tools. Microsoft’s own AI readiness guidance, announced November 17, 2026, focuses on “modernizing authentication and deploying monthly updates.” IBM’s 2026 Cost of a Data Breach Report found that the most common causes of breaches targeting AI were compromised APIs, apps or plug-ins and cloud misconfigurations, at 27% each. Cisco’s 2025 AI Readiness Index found that only 15% of organizations have networks fully ready for AI.
2. Operational performance optimization (monitoring, technology lifecycle management). 24/7 monitoring, change management, and a planned refresh of aging equipment, old keys and stale accounts, so routine changes like Microsoft’s and the FAA’s don’t cascade.
3. Security (cybersecurity, vulnerability management). Assessment, prevention, response and governance, including vendor and AI governance, across IT and physical security systems, from servers to cameras and door controllers. Verizon’s 2026 Data Breach Investigations Report found that vulnerability exploitation is now the top way into a breach, at 31%, and that a third party was involved in 48% of breaches. Microsoft’s November 2026 research found that for 2026 vulnerabilities, the median time from disclosure to CISA’s exploited list was 3 days.
Enterprise-grade controls at small and medium business prices. Most smaller and mid-sized companies assume the tools used by federal agencies, major financial institutions and global enterprises are out of reach. As part of a comprehensive solution, BTI deploys those same enterprise-grade tools from our high-security enterprise partners at small and medium business prices, managed under one roof. Here is what those layers are designed to give your business:
- Governed cybersecurity controls. Clear policies, documented reviews and auditable activity logs across users, devices and cloud apps, so security isn’t ad hoc and you can show what controls were in place.
- Protection level agreements for patching and maintenance. Defined schedules, explicit patch testing and whitelisting, pilot rollouts and post-patch verification across servers, endpoints, cloud apps and embedded devices. Because BTI explicitly tests and whitelists patches as part of our routine patching maintenance for clients, critical fixes like this week’s nginx, vCenter and Java updates move quickly without treating your production systems as a vendor’s test lab.
- Zero trust access. Identity-based verification for every user, device and application, replacing broad legacy VPN access so a single stolen login doesn’t open the whole network.
- Data loss protection (DLP). Cloud-enforced controls that watch for sensitive customer, financial and health data, including text inside images, file uploads and web forms, and help stop it from leaving where it shouldn’t.
- AI governance. Automatic discovery of which AI tools and cloud apps staff actually use, policy limits on what data can be shared with them, and logging that matches what regulators and frameworks expect.
Those are the exact controls the privacy and compliance rules above describe, from access controls and user activity logging in the Safeguards Rule to documented limits on what AI tools can reach. No tool makes a business compliant on its own, and no toolset eliminates every risk, but governed controls and clear records reduce exposure and help you demonstrate reasonable care.
Whatever is missing, whether it’s an unpatched vCenter, a camera server running old PHP or an AI tool with too much access, gets the same treatment: spot it, assess it, evaluate it, fix it. Some businesses need one tool, some need a partner alongside their team, and some want full management.
When leadership change is occurring even at the top of U.S. cybersecurity after a 10-day wave of disruptions, the inference for every business is clear: the old way of handling cybersecurity and data loss prevention, simply outsourcing standalone tools and SOC services while leaving daily IT operations untouched, can’t succeed. Real protection requires an overhaul of IT operations, including vigilant co-managed patching and ongoing vulnerability research and remediation, paired with state-of-the-art tools for zero trust, data loss prevention, and AI governance and controls.
Why BTI
That operational foundation is what BTI Communications Group has delivered since 1985. We were founded in Chicago that year. Today we have offices in Santa Fe Springs, Downers Grove and Phoenix. We deliver an AI-enabled a la carte, co-managed, or fully managed toolset. It is customized for each client’s IT infrastructure and support capabilities. We deliver it fast and at the scale our clients need. Month-to-month managed IT is available. Every BTI support agreement includes SLAs with 4-hour onsite emergency response. 24/7 monitoring runs through high-security enterprise partners managed under BTI’s SLA. BTI designs, sells, installs and commissions it all, from networks, cloud and AI to cybersecurity, data governance and connectivity. That includes access control, alarm and camera systems that feed your security operations and business intelligence.
Want to know where your gaps are?
Frequently asked questions
What does a third-party data breach mean?
A third-party data breach happens when attackers compromise an outside vendor, contractor or software provider that holds or connects to your data, rather than breaking directly into your own network. Even when your own systems stay untouched, exposed names, emails or account records still create phishing, fraud and compliance risk for your business.
What should I do when my vendor has a data breach?
Ask the vendor in writing what data was exposed, when the access occurred and whether any shared logins or integrations were affected. Rotate any shared credentials or keys, alert staff to watch for targeted phishing tied to the leaked details, and check with your attorney on state or industry breach notification duties.
What are the risks of shadow AI?
Shadow AI is any AI tool, browser extension or code library staff use without IT approval or oversight. Without an inventory and policy controls, sensitive customer, financial or health data can leave through unmonitored prompts and uploads, and unvetted AI tools or gateways can introduce security flaws and stolen credentials.
Were the banks hacked?
According to the companies, no. JPMorganChase, Mastercard and Visa each said the incidents happened at third-party providers, that their own systems and payment networks weren’t affected, and that card numbers weren’t involved.
Were the Microsoft, FAA and NSA outages cyberattacks?
No. Microsoft called its outage “a technical issue not related to cyber or AI / agents,” the FAA says it has no indication of a cyberattack, and Director of National Intelligence Jay Clayton confirmed the NSA outage came from urgent PHP patching rather than an attack. All three followed routine maintenance or updates. The USDA outage is different: USDA took applications offline to reset database passwords after uncovering a security incident.
What is AI-ready infrastructure?
It’s a network, identity setup, data and security that are ready for AI tools. That means enough capacity and redundancy, MFA and modern authentication, organized and protected data, and governed, monitored access for every AI tool and agent.
How many states have comprehensive privacy laws?
Depending on the tracker, 23 to 24 states have passed one (IAPP lists 23, with 19 in effect today, while MultiState counts 24). The rest take effect between January 1, 2027 and January 1, 2028. Every state also has a breach notification law.
Do the new AI safety proposals apply to my business?
Mostly not directly, yet. California’s November 18 recommendations aren’t law and mainly target large frontier AI developers. The UK guidance is written for financial firms. But guidance like this often shapes what regulators, courts and plaintiffs’ lawyers treat as reasonable care, so it’s worth following. This isn’t legal advice; ask your attorney about your own obligations.
Does a security tool make my business compliant?
No. Compliance depends on your policies, records and how you run your business. The right tools, such as access controls, data loss prevention, MFA and activity logging, give you the controls many of these rules expect and the records to show you followed them.
Can an IT provider prevent every breach or outage?
No. No toolset stops a breach at someone else’s vendor or an outage inside Microsoft. A good partner makes sure your own gaps are found and fixed first, so incidents are smaller, shorter and less disruptive.
