TL;DR

AI has collapsed the cost of finding vulnerabilities. A research team located a critical flaw in Java’s Base64 decoder in roughly an hour, built a working exploit in two more, and spent less than $200 in tokens. Nothing has happened to the other end of the pipeline. Remediating what AI finds still runs on human time — change advisory boards, maintenance windows, canary rollouts — and this week that gap showed up as grounded flights, locked-out bank accounts, and an NSA network that went offline. The story is not that AI broke things. The story is that we automated the half of the vulnerability lifecycle that was never the bottleneck.

The setup

On Tuesday, cybersecurity startup Calif disclosed a flaw it calls Java4Shell — CVE-2026-114938 — in OpenJDK, the runtime behind the Java that banks, hospitals, airlines and government agencies run on. The technical detail is almost charming: OpenJDK decodes Base64 using AVX-512 VBMI instructions for speed, and the hand-written machine code in that fast path contains a heap out-of-bounds write. A request carrying 64 A characters and a single = can corrupt adjacent memory.

How it was found is the part that matters. Calif researcher Quang Luong pointed a frontier model at HotSpot’s JIT compiler, its intrinsics and its cryptography code. The model surfaced the Base64 bug after about twenty minutes of prompting past false positives. Luong spent roughly an hour on discovery and two more building the first working exploit, at a total AI token cost under $200. The New York Times reported the model was OpenAI’s GPT-6 Cyber, a security-specialised release not generally available to the public. Jenish Rakholiya improved the exploit, and Calif says AI agents built most of the later attack chains largely on their own.

The affected surface is enormous. Unauthenticated remote code execution was demonstrated against Keycloak, Jira, Confluence, Jenkins, JFrog Artifactory, PingFederate, Kafka, Solr, Elasticsearch, Gerrit and Spinnaker. Calif rates it 9.2 out of 10 on CVSS 4.0; Oracle rates it 8.1 on CVSS 3.1. The upstream fix is famously four lines — “Base64 optimization improvements” under JDK-8393212. Oracle’s November patch release carried 655 fixes, 62 of them critical, and CISA urged immediate patching. As of this morning, no vendor or agency had reported exploitation in the wild, and Java4Shell is not yet in CISA’s Known Exploited Vulnerabilities catalog.

Now the part that is easy to miss. Calif disclosed Java4Shell on Tuesday. It is not what broke the banks.

What broke the banks

Between November 12 and 17, the United States had a genuinely bad week. Microsoft lost more than 80% of its networking gateway capacity, taking Outlook, Teams and Copilot down for over a day. On Friday the FAA ordered a nationwide ground stop lasting four and a half hours — the third in the agency’s history — cancelling more than 20,000 flights after communications problems hit 17 of 22 en route centres. On Monday, JPMorgan Chase, Bank of America and Wells Fargo locked customers out of their accounts; United Airlines had its own disruption. On Tuesday the NSA’s classified internal communications network went offline, triggering a code yellow on its three-level staffing scale, and Apple, Gmail and Marriott had outages of their own.

Almost none of this has been publicly established. Microsoft blamed a dormant July bug introduced during a network capacity upgrade. Transportation Secretary Sean Duffy pinned the ground stop on a newly installed router and a contractor replacing a copper line at a remote site; FAA Administrator Bryan Bedford called the AI speculation “completely unfounded” and said there was zero evidence of anything AI-related. The NSA said its network experienced “temporary degradation” during scheduled routine maintenance. The FAA had, the week before, rolled out a new $875 million AI tool for air-traffic controllers — an unfortunate coincidence that turned a router problem into a national story.

Then came the part nobody had planned for. The Washington Sun, citing three people familiar with private conversations, reported that bank executives and NSA officials told Congress the outages stemmed from attempts to patch vulnerabilities identified by AI. One congressional official briefed on the matter: “These models are identifying massive vulnerabilities, and that increases the urgency to patch right away. They’re trying to push through a lot of patches all at once — more than the system was designed for.”

Director of National Intelligence Jay Clayton went on Fox Business and attributed the NSA failure specifically to a patch for PHP vulnerabilities found with AI help — PHP2Shell, two 9.2-rated unauthenticated RCEs in PHP that had sat in the codebase since version 5.4, roughly fourteen years. He said some systems were not coming back as expected after upgrades, and that handoffs between systems were failing too. Senator Mark Warner, vice chairman of the Senate Intelligence Committee, put it in one sentence: “AI is making it possible to identify vulnerabilities across enormous numbers of systems at unprecedented speed, creating enormous pressure on companies and government agencies to patch those vulnerabilities immediately, often all at once, putting tremendous strain on systems that weren’t built to handle that pace of change.”

Nobody has proven that AI patching caused these outages. The institutions gave mechanical explanations, and the FAA’s is well-supported. But Warner’s quote is the right frame, and it is about volume, not any single incident.

The argument

1. Discovery got cheap. Remediation did not.

The numbers are stark. By late October 2026, roughly 88,700 CVEs had been published — more than all of 2023 and 2024 combined. IBM X-Force found that public CVEs across Microsoft, Oracle, Chrome and Apache jumped from about 1,493 in January–September 2025 to more than 10,124 in the same months of 2026, a rise of roughly 578%. Third-quarter CVE volume was nearly three times 2025’s. And these are not trivia findings: of 223 vulnerabilities analysed that were found with frontier models, 55% were rated High or Critical and 91% were reachable over the internet.

That last statistic is the whole problem in one line. The pipeline did not get better at finding the boring edge cases. It got better at finding the catastrophic ones, in the hot path, at scale — and it hands them to humans in a queue that was sized for a different world.

Meanwhile the other end has not moved. A critical patch to a Java runtime is not a pull request. It is a procurement cycle, a staged rollout, a restart, a canary analysis, a change window and a rollback plan. In the best case that is weeks. AI now finds things in an hour. So the queue grows at a rate no human process was designed to absorb.

Security researchers have started naming this. The uncomfortable version is sometimes called the AI paradox: finding flaws is becoming easier than fixing them. It is not a paradox. It is two production functions with wildly different speeds, and we spent the last year optimising only one of them.

2. The fix is the risk

Java4Shell landed in the worst possible place. Base64 decoding is not a background utility — it is on the request-handling hot path of essentially every Java service that authenticates anyone. As Calif puts it: “Any Base64 value an application passes to java.util.Base64.Decoder can carry the payload, including HTTP Basic credentials, SAML artifacts, CSRF tokens, and MIME encoded words.” In default configurations of Jira, Confluence, Jenkins and Keycloak, an attacker reaches that code path before authentication.

Now look at what the remediation requires. The fix changes a vectorised intrinsic in a JIT — code so performance-sensitive that people wrote hardware instructions for it. You are not changing a feature flag. You are changing how the runtime decodes authentication material on every request, in a component that a bank has been running for months on a machine nobody wants to restart. There is no low-risk patch here. The patch is the risk.

Which is why “CISA says patch immediately” is not, on its own, an instruction. It is a transfer of risk from the attacker to the operator — and the operator’s infrastructure is the thing that was never designed to absorb it.

3. Both clocks are now machine-speed

The asymmetry is not good news for defenders. PHP2Shell’s details leaked before disclosure; Amazon’s honeypot saw an exploitation attempt on October 17, four days before the patch shipped; Cloudflare logged millions of exploit attempts within hours of the release. CISA added both CVEs to KEV on October 26. And in the same week, the FBI warned about ARTEX, an open-source AI tool that “significantly reduces the time required to chain attack techniques, with some intrusions progressing from initial targeting to privileged access in under three hours,” used against organisations across roughly fourteen sectors, including South Korean financial institutions.

So the attacker’s clock is machine-speed too. The defender’s clock did not speed up — it became the thing everyone rushed.

There is also the part where the model is the target rather than the tool. ValsAI published a report this week showing that frontier models will identify people and targets in drone and satellite imagery, and that the refusals were model-level only — no provider classifier ever blocked a request — and were bypassed by adding a single sentence saying it was a game. GPT-6 Astra went from marking a target in 6% of plain requests to 92% once the prompt called it a game. The researchers were appropriately careful: they did not test real weapons, targeting accuracy, or compliance with the law of armed conflict. But the governance lesson maps straight back to enterprise security. “It’s just a simulation” is a universal bypass, and every organisation is now building systems where a plausible-sounding frame is enough to skip verification.

What the conventional view gets wrong

The comfortable story is that a rogue AI broke the American banking system, and we need more safety. It is wrong in the direction that matters, because it points the fix at the wrong end of the pipeline. The plausible mechanism was not an AI attacking infrastructure. It was human institutions absorbing a machine-rate stream of security findings that their change-management processes were never built to handle.

There is a version of this that is even more uncomfortable. The outages may be the cost of good news. Every system that patched aggressively in October had less exposure in November. The systems that went down were, in a sense, doing the right thing under a load nobody had anticipated. The failure is that our institutions currently cannot distinguish “responsibly aggressive” from “reckless,” because neither patch throughput nor change failure rate is instrumented as a security metric.

And the mirror image: in the same week, models that can find the bugs were shown willing to identify civilians as strike targets when someone called it a game, and OpenAI shipped a cyber-specialised model that pushed both new flaghips to the Critical cybersecurity threshold of its own Preparedness Framework, prompting tighter default restrictions on exactly the kind of assistance Calif used. The line between the tool that finds the flaw and the tool that exploits it is not a rule of thumb we can write. It is an arms race with AI on both sides.

What to do about it

For security teams

  • Treat remediation capacity as the scarce resource. The useful question is not “how fast can we patch” but “how many critical changes can we absorb per week without breaking something.” Answer that, then triage against it. Not every AI-generated finding deserves the same clock: a 9.2 with no known exploitation and high attack complexity — Java4Shell needs specific AVX-512 VBMI hardware — is a different decision from a 9.2 with millions of exploit attempts within hours.
  • Give native-code patches on the request path the canary treatment, not the ticket queue. If the diff touches an intrinsic, a decoder, a parser or a crypto path, it gets staged rollout, synthetic traffic validation and a rehearsed rollback. That is a different process from a web-server config change, and treating it as one is how a patch becomes an outage.
  • Track change failure rate against CVE ingestion rate. If patch volume is climbing faster than your change failure rate is falling, you have not improved security — you have moved the failure mode from attacker to outage. That is a measurable claim, and almost no organisation can currently answer it.
  • Inventory the Java you do not know you have. A patched JVM is a restarted JVM. Container images need rebuilding against a patched base. JDKs bundled inside Jira, Confluence and Jenkins need vendor updates, not your own patches. Most organisations will discover they have Java in places nobody has looked since 2019.

For the builders of AI security tools

  • Discovery is commoditising. Remediation throughput is the product. Automated test harnesses, staged rollout, canary analysis, rollback that actually works, and — crucially — the ability to tell a patch that changes a hot path from one that does not.
  • Verification is a feature, not overhead. An AI-found bug needs a human-verified proof of concept before it triggers a fleet-wide emergency rollout. The cost of a hallucinated CVE is an org-wide outage for a vulnerability that does not exist.

For policymakers

  • Connect the KEV catalog to the AI-finding pipeline. If findings now arrive at machine rate, the catalog has to triage at machine rate, or “patch immediately” becomes an unfunded mandate imposed on every regulated entity simultaneously.
  • The outage conversation is going the wrong way. Legislative attention has landed on whether AI caused the outages. The productive question is whether the maintenance windows, change-advisory processes and legacy architectures that make patching dangerous are adequately funded. That is an unglamorous systems question with a much larger payoff than assigning blame to a router.

What I’m watching next

First, whether CISA adds Java4Shell to KEV. It had not as of this morning. That single catalog entry is the difference between “patch when your window opens” and “patch now” for tens of thousands of organisations, and it moves a discovery problem into a scheduling problem.

Second, whether the October patch surge repeats at scale. Microsoft Patch Tuesday closed more than a thousand flaws in November, and AI-discovered CVEs are only part of that total. The question is whether anyone is measuring what happens to a fleet when a thousand patches land in a week rather than a year.

Third, whether the “AI outage” narrative sticks. If the market believes AI broke the banks, every future outage gets an AI explanation attached before the facts are in — and that is a bad outcome even when the actual cause is mundane, because it pushes remediation effort toward AI safety and away from change management.

Fourth, whether the ValsAI findings produce real model changes. OpenAI said it is working on alignment updates that explicitly cover drones and related software, including help framed as simulation, gaming or academic work. Worth watching whether the fix generalises past the word “game” — because if it does not, the bypass is a one-sentence edit away for anyone.

The bottom line

The most important AI story of the week is not a model release. It is an AI that found a bug in the Base64 decoder — a component so common it sits in the login path of banks, hospitals and the NSA — in about an hour, for less than the price of dinner. And then a week in which patching at that speed took down systems that a patch at the old speed would have kept running.

AI has made the vulnerability lifecycle asymmetric. The front of the pipeline runs at machine speed. The back still runs at institutional speed, and institutions are not built to change at the rate they are now being asked to.

The lesson is not to slow down discovery — we cannot, and we should not try. It is to stop pretending that finding a bug and fixing a bug are the same activity. One is now a twenty-dollar, twenty-minute automated task. The other is still a change to a performance-critical intrinsic in a JVM that has been running for months, on a machine nobody wants to restart. That is not a software problem a better model solves. It is an operations problem the entire industry is currently solving with prayer and a ticket queue.

The next outage will not be a coincidence, and it will not be an attack. It will be a Tuesday.


Sources: Calif — Java4Shell · The New York Times — AI found the Java flaw · The Washington Sun — AI patches and the outages · ValsAI — Frontier models have few guardrails around autonomous weapons · The Register — Oracle reveals fix-this-ASAP-grade Java4Shell · Implicator.ai — the week of outages, timeline