In the last few weeks, flights stopped, bank customers lost access to their money, and the NSA’s classified network went dark. Officials say none of it was an attack. What they describe is a patching system pushed past its limits: AI is finding vulnerabilities faster than organizations can deploy fixes without breaking what already works.
The week everything broke
Between November 13 and 18, four of the country’s most critical systems failed. The FAA halted every departing flight in the United States for more than four hours, the longest nationwide ground stop since September 11, 2001. More than 9,100 flights were canceled that day. Microsoft 365 went down for nearly 24 hours. Customers at some major banks couldn’t reach their accounts. The NSA’s classified internal network went dark.
The Washington Sun reported that bank executives and the NSA told members of Congress the disruptions came from rushed fixes for vulnerabilities identified by AI. “They’re trying to push through a lot of patches all at once, more than the system was designed for,” one congressional official said. Palo Alto Networks CEO Nikesh Arora was more direct: “Some of the recent outages are a consequence of a deluge of patches.”
The volume explains the pressure. On November 13, the count of published vulnerabilities for 2026 passed 100,000, more than double all of 2025. That count covers only disclosed vulnerabilities. It leaves out findings from frontier models like Mythos that are still under embargo. Open-source models are catching up to frontier models quickly, giving attackers similar capabilities, and the median time from disclosure to exploitation has dropped from 120 days to 80. Every security team is being told to patch faster.
The US Government is already responding. A bipartisan bill introduced this week would create a federal board to investigate outages caused by faulty updates, even when no attacker is involved. Cybersecurity journalist Nicole Perlroth called the week “just the opening chapter.”
AI has made finding vulnerabilities fast and cheap. The systems and teams that code and deploy fixes were never built for this volume, and every fix that lands as a breaking change is another chance to take production down.
Changing dependencies has always been a challenge
Anyone who has run software in production knows that changing a dependency is never just one change. A new release carries the fix you need along with new APIs, removed features, new transitive dependencies and different build settings.
That’s why enterprises pin versions. They upgrade on their own schedule, after testing, and treat every dependency change as a risk to manage.
Most vulnerability fixes arrive as exactly that kind of upgrade. Maintainers fix the flaw in a new release, and a team on an older version has to move forward to use it.
A team can test and absorb a handful of updates a quarter. AI-driven discovery now asks security teams, collectively, to triage thousands of updates per week, and that is where systems start to break.
Open-source Python is particularly hard
Python runs the models, analytics and data pipelines behind most of modern finance, research and AI. Much of it rests on community projects kept alive by small teams of volunteer maintainers, who ship fixes in new releases on their own schedule.
Many of the most important packages are thin Python layers over compiled C, C++ and Fortran. And each package ships as wheels and conda packages across dozens of combinations of operating system, CPU architecture and Python version, every one of them a separately built artifact. A fix that works on one build can fail, or quietly behave differently, on the next.
One reason this problem is so tricky: a numerical library can be patched and rebuilt with a slightly different compiler or math library, install cleanly, pass its test suite and still return different numbers. In a risk model or pricing engine, that’s a silent change to business results that no alert will catch. A firm could be trading on mispriced positions or misreported risk for weeks before anyone notices.
Multiply that across thousands of findings a week and you get what we saw this month: a system overloaded and breaking under its own fixes.
What a safe fix looks like
The alternative to an upgrade is a drop-in replacement: the exact package a team already runs, with only the vulnerability fixed. A safe one has four properties.
- The exact version. The fix goes onto the version running in production, so the application around it doesn’t change.
- The original build. Compiled packages are rebuilt with the same compiler, flags and linked libraries as the original, so they behave the same way.
- Expert verification. Someone who knows the package confirms the fix, closes the flaw and changes nothing else, while documenting exactly what changed and what to test.
- A path upstream. Once any disclosure embargo ends, the fix goes back to the open-source project, so the whole ecosystem benefits.
What teams can do now
Few teams can build every fix this way on their own. Every team can lower the odds that a security fix becomes the next outage.
- Know what you run. Inventory the pinned and compiled packages in production, and which applications depend on each.
- Ask what a fix changes. Before deploying, find out whether it’s a backport to your version or an upgrade that brings other changes with it.
- Compare outputs after a rebuild. For numerical packages, run real workloads against the old and new builds and check the results match. A passing test suite isn’t enough.
- Stage every rollout. Deploy fixes in waves and keep a tested path back to the previous version.
What more than a decade of building and patching Python at scale has taught us
Anaconda has built and distributed open-source Python for more than a decade, across Linux, macOS and Windows, on x86 and Arm. Our packages were downloaded 1.1 billion times in September alone. The main lesson from that work is that the patch is usually the easy part. Most of the effort goes into rebuilding a package so that nothing else changes, and proving it before deploying a fix that will bring a system down.
Development teams are struggling to find, fix and rebuild their software. Security teams are being handed thousands of fixes and told to ship all of them, and every fix that arrives as an upgrade is a chance to break production. What the industry needs now are fixes delivered at speed and volume that are drop-in replacements for what teams already run, built the way the original was built. That’s how organizations clear this dramatically growing volume without breaking what’s already in production.
That experience is why we’ve started working directly with a small number of enterprises on this problem. We deliver private, expert-reviewed fixes on the exact versions of open-source Python they run, including for vulnerabilities that haven’t been disclosed yet.
Nearly every organization now faces this AI-powered challenge, where resiliency is critical. The organizations that successfully move through this period will be the ones that can fix everything AI finds without breaking what already works.
If you have this challenge, contact us to explore working with the Anaconda team.