Planes Stopped and Everyone Blamed AI. It Wasn't AI. Next Time, It Might Be. Anthropic's Frontier Red Team said Monday it will cut the disclosure window for high- and critical-severity vulnerabilities its AI finds from 90 days to 30, starting with reports filed on or after Jan. 1, 2027, because AI models are increasingly likely to find the same flaw another model turned up. The announcement followed a week in which AI was wrongly suspected in major failures: the FAA's nationwide ground stop, which lasted more than four hours and contributed to more than 14,000 canceled flights across Friday and Saturday, was caused by faulty equipment, not the agency's $875 million SMART traffic planning tool, FAA Administrator Bryan Bedford said, and Transportation Secretary Sean Duffy said SMART "isn't plugged in." A Mississippi River tanker collision in New Orleans that same day was traced by federal investigators to a rudder that stuck two minutes before impact, not AI. Anthropic said Monday it will give software makers less time to fix serious security flaws its AI finds before it goes public with them. For high- and critical-severity bugs, the window drops from 90 days to 30, starting with reports filed on or after Jan. 1, 2027. In a post from its Frontier Red Team https://www.anthropic.com/research/coordinated-vulnerability-disclosure-timelines , the company said the old timeline assumed fixes would happen at human speed and that nobody else was likely to find the same bug in the meantime. It no longer believes either. As AI models get better at finding vulnerabilities, a flaw one model turns up is increasingly likely to be found by another. The post landed at the end of a week when the first question about almost every failure was whether AI caused it. We think that instinct says more about how AI is being deployed than about any one outage. Companies and governments are wiring AI into systems that weren't built for it, often faster than they can explain what the new tool touches. When something breaks, nobody can rule the AI out quickly, so people assume it might be involved. Friday's suspects On Friday afternoon the FAA stopped every departing flight in the country. Four days earlier, the agency had expanded SMART, its $875 million AI traffic planning tool, to all 22 of its en route centers. Speculation that the tool was to blame spread online almost immediately, KCRA reported https://www.kcra.com/article/faa-ground-stop-raises-questions-after-rollout-of-new-ai-air-traffic-system/74428995 . By that evening, Homeland Security Secretary Markwayne Mullin was telling reporters at the White House that DHS was "looking at that," The Well News reported https://www.thewellnews.com/homeland-security/dhs-chief-says-faa-outage-shows-no-sign-of-attack/ . It wasn't SMART. FAA Administrator Bryan Bedford said Saturday that faulty equipment caused the ground stop and that investigators had found zero evidence of anything AI-related, ABC7 reports https://abc7.com/live-updates/faa-ground-stop-live-updates-what-know-nationwide-flights-are-grounded/20038887/ . Transportation Secretary Sean Duffy said SMART "isn't plugged in," according to Quartz https://qz.com/faa-smart-ai-air-traffic-scrutiny-ground-stop-111626 , citing The Hill. Air Space Intelligence, the startup that built SMART, said its products had nothing to do with the outage, per WISN https://www.wisn.com/article/questions-faas-new-smart-ai-system-ground-stop/74439726 . The ground stop lasted more than four hours, AVweb reports https://avweb.com/aviation-news/faa-nationwide-ground-stop-lifts-smart-ai/ , and more than 14,000 flights were canceled across Friday and Saturday. The same day, an asphalt tanker lost control on the Mississippi River and rammed a docked cruise ship in New Orleans. Everyone aboard was accounted for and no one died, NBC News reports https://www.nbcnews.com/news/us-news/rescue-underway-oil-tanker-collides-new-orleans-river-boat-rcna609235 . Federal investigators now say the tanker's rudder got stuck two minutes before impact, according to Splash247 https://splash247.com/steering-failure-blamed-for-new-orleans-tanker-cruise-ship-collision/ . Microsoft users had also spent part of the week offline, and Mullin made a point of saying the FAA problem wasn't connected to that outage, per CNN https://www.cnn.com/2026/11/13/us/live-news/flights-ground-stop-faa . Neither failure has anything to do with AI, as far as investigators can tell. One came down to old equipment and the other to a rudder that wouldn't turn. We were jumpy too The State of AI was part of this. On Friday we ran a story headlined "Feds Are Now Looking at Whether the FAA's New AI Tool Caused Today's Nationwide Ground Stop." The reporting was accurate, and we said plainly that nobody had tied SMART to the outage. The headline still pointed readers toward an answer that turned out to be wrong. The reflex wasn't irrational. SMART went from three airports to the whole country in about seven weeks. The air traffic controllers' union said it had no role in designing or testing the tool, NPR reported https://www.npr.org/2026/09/21/nx-s1-5976816/faa-ai-manage-airspace at launch, and House Democrats had asked for an audit of the software last month. When a new system goes live inside an old one and the old one breaks days later, asking about the new system is fair. The problem was that the question became the story within hours, while the answer took about a day. Why next time could be different Anthropic's post describes a world where AI finds security flaws faster than the people responsible for them can ship fixes, and where the same flaw is increasingly likely to be found more than once. Anthropic uses that to argue vendors should patch faster. It's also a reminder that the people trying to protect these systems aren't the only ones with access to capable models. The post also makes an exception that matters here. Software that's hard to deploy, including critical infrastructure, may get a longer timeline negotiated with the vendor. That describes the FAA's world well. The last nationwide ground stop, in January 2023, was traced to a damaged file https://en.wikipedia.org/wiki/2023 FAA system outage in a decades-old system. This one came down to faulty equipment. Systems like these take months or years to change, and they now sit in a world where AI is getting steadily better at finding their weak spots. The AI theory was wrong on Friday. With more AI going into old systems, and AI getting better at finding the flaws in them, it won't always be wrong. Our take Our view is that the next outage blamed on AI may well be caused by AI, and most companies won't be able to confirm it or rule it out quickly when it happens. Leaders should put a direct question to their teams. If a critical system went down this afternoon, could you tell within an hour whether your AI played a part? On Friday it took the FAA, DHS and the company that built SMART roughly a day to clear the tool, and the speculation ran ahead of them the whole time. Most companies won't have that many people working the problem. Getting to a fast answer takes work before launch. Teams need to know which systems their AI connects to and what it's allowed to change, and they need records of what it actually did. They also need the people who run the older systems involved from the start, which is the step the FAA skipped with its controllers. Anthropic's post raises the stakes. If flaws are being found faster and by more people, the legacy systems underneath your AI rollouts are more exposed than they used to be. Being able to rule AI in or out quickly should be a launch requirement, reviewed as seriously as security. SMART was a bystander on Friday. The companies that handle the first real AI-caused outage well will be the ones that did this work while their AI was still innocent.