No attack has been found behind the six US outages of Nov. 12 to 17: by their own accounts, five of the six operators broke their own systems, the FAA still cannot say what caused its failure, and the one confirmed attack that week came through Optimizely, a tool big companies run on their websites. The six were Microsoft's business email, US air traffic control, several big banks, United Airlines, Apple and the NSA's classified network. Officials ruled out an attack before the FAA knew what broke, and the Optimizely attack was found by an outside researcher, not by the 178 organizations serving it to their visitors. The danger this fall was not AI going rogue. It was that the people running critical systems could not tell a bug from an attack, while AI sped up both the discovery of software flaws, which forces rushed fixes, and the attacks that use them. This edition takes the hard questions one at a time.
Sponsor
Arca pays for your AI sessions
Got hundreds of Claude Code, Codex or Cursor sessions just sitting there? Might as well get paid for them.Is AI going rogue?
Not in the way people fear. No company or agency that failed last week has blamed AI, and officials say they found no attack of any kind. The risk in our reporting is people deploying AI that is fragile and easy to trick faster than they can secure or govern it.
The evidence is in three places below. Attackers are running AI agents, programs that pursue a goal on their own, against hundreds of organizations, with a person choosing the targets. AI models in tests picked a school as a military target, and a camera-image attack steered a self-driving model into crashes while its explanation stayed the same. And AI labs' own test agents reached real systems they had no permission to touch; on Nov. 16 California sued OpenAI over AI models that escaped their test environment.
How did Washington rule out an attack before it knew the cause?
When the FAA stopped every departure in the country on Nov. 13, Washington first treated it as a possible attack. Fortune reported that the White House Situation Room was activated and the FBI was on site. By the next day FAA Administrator Bryan Bedford said, "I think we've all concluded that it's not a cyber issue." On the same day Transportation Secretary Sean Duffy said the FAA was still trying to work out whether its contractor, L3Harris, had made a mistake or installed a faulty router.
What is known is the mechanism. Equipment at two FAA sites, in Amarillo, Texas, and El Dorado, Ark., flooded the network with more than 1,000 connection attempts every four to five seconds. Duffy told CBS's Face the Nation, "We think it was the installation or the router that was the bad actor."
That is a suspect, not a cause. On Nov. 17 Duffy said engineers could recreate the El Dorado failure but still did not know why it happened. Five days after the ground stop, nobody has explained why both sites failed at the same moment.
Homeland Security Secretary Markwayne Mullin's language firmed up over the same days, without new public evidence. Right after the ground stop he said, "there's no chatter, we haven't seen any footprints."
On Nov. 16, Fortune reported, he told Fox News he had "98% certainty" there was no attack, and relayed a theory from a tech contact that the cause was SI, the administration's term for superintelligence, overwhelming older systems.
On Nov. 17 Mullin said he had "complete certainty that we are not under attack."
Lawmakers asked for the basis. Sen. Josh Hawley, a Missouri Republican, said that if the failures were unconnected it was "the greatest coincidence in the history of the world", and that the public should not be kept in the dark.
Mullin testified before the House Homeland Security Committee on Nov. 18, as this edition was going out. The question he owed the committee an answer to: what evidence lies behind "complete certainty"? We will report what evidence he gave.
Six critical systems failed in six days. If not hackers, then what?
By their own accounts, the operators broke their own systems:
- Microsoft business email, Nov. 12 to 13. Exchange Online was down for about 24 hours because of a defect in a July update that Microsoft's load tests missed.
- FAA, Nov. 13. Equipment at two sites flooded the network. Cause unknown.
- Banks, Nov. 16. Customers reported problems at Chase, Bank of America, U.S. Bank, Capital One, Truist, Wells Fargo and Zelle. Chase cited a technical issue. U.S. Bank acknowledged delays on some Zelle payments to other banks.
- United Airlines, Nov. 16. About 90 minutes of departure delays from what United called a technology issue.
- Apple, Nov. 17. iCloud sign-in, the App Store and Apple Pay failed after an internal update.
- NSA classified network, Nov. 17. Director of National Intelligence Jay Clayton said "this is not a cyber attack."
The NSA case is where AI enters, and it is not as a rogue actor. CNN reported that the network went down while the NSA was patching PHP2Shell, two serious flaws in PHP, the programming language behind a large share of websites. The security firm Wiz, one of three teams that reported PHP2Shell, says it found the flaws with help from AI models. Once a flaw like that is known, everyone who runs the software has to patch fast, because attackers move fast. Amazon's decoy servers recorded an attempt to exploit PHP2Shell four days before the fix was released.
Rushed patches on old systems break things. The former government official James Lewis told NBC: "When people try to patch it, they break it."
Mike Chapple, who teaches IT at Notre Dame, told CBS the failures together point to "the complexity and fragility" of the systems.
That is the accountability question. Microsoft owes a final report on Nov. 20. The FAA and L3Harris owe a root cause.
Why did an outside researcher find the one real attack, not the 178 organizations serving it?
Optimizely sells a tool companies load on their websites to test different versions of a page. On Nov. 16 an attacker planted code in it, and for most of that day the tool served the code to visitors of 178 organizations, BleepingComputer reported, including CNN. The code could swap crypto wallet addresses, skim card details and deliver iPhone exploit kits.
The attack was uncovered by Scott Helme, a security researcher whose company, Report URI, checks the scripts websites load. None of the affected organizations has said it caught the code itself. Optimizely told BleepingComputer the attacker used a flaw in its account signup flow, which it has closed. That is the weak point the outages distracted from: a big company's website runs code written and controlled by other companies, and most never check what that code does on the day it changes. Optimizely still owes an account of how a signup flaw gave an attacker write access to so many customers' code.
Is AI arming attackers faster than defenders?
Faster, yes. Better, not yet clearly. ARTEX is a free tool, posted on GitHub in July, that coordinates AI agents to find weaknesses in a network and break in. It was built for penetration testing, hacking done with a company's permission. By Nov. 12, Cloudflare said ARTEX and copies of it made up more than 60% of the AI hacking tools it could identify, and daily traffic from such tools had grown 30 times since Oct. 1.
The security firm CrowdStrike said VAPOR PANDA, a group it ties to China, used ARTEX with DeepSeek, a Chinese AI model, against more than 200 organizations in about 14 countries starting Oct. 17. It confirmed the attackers got further inside nine of them, and that a human operator directed the campaign.
In Seoul, police investigated a gang behind breaches of more than 90 organizations and 8.2 million records, most carried out with ARTEX.
The FBI warned on Nov. 13 that some ARTEX break-ins went from first contact to administrator access in under three hours.
The tools people use to build with AI were targets too. Hackers who broke into Fireworks AI, a company that runs AI models for other businesses, used what they found to publish two fake versions of the Claude Code add-on for Cursor, a popular AI coding app. Cursor estimates up to 5,000 users may have installed them.
On defense, AI is finding flaws faster than people can fix them, and the deadlines are shrinking. Anthropic said it will give software makers 30 days instead of 90 to fix the serious flaws it reports, because others increasingly find the same flaws on their own.
Shorter deadlines mean more rushed patches, the same pressure that hit the NSA. The sceptic's case also deserves a hearing. Nathan Lambert, who writes the Interconnects newsletter, argues that warnings about AI cyber risk have run ahead of the harm documented.
On the evidence so far, AI gives attackers speed and scale, with a person still in charge.
What did the experts get right and wrong in real time?
AI Weekly follows 2,120 AI researchers, engineers, journalists and commentators on Bluesky and X. From Nov. 12 to 18, 224 of them posted 959 times about the outages. We used Claude to sort each post by the cause it pointed to, then checked the results. Of the posts that named a cause:
- 66% blamed fragile systems, staff cuts or bad updates.
- 27% blamed AI in some form.
- 7% suspected an attack. That share peaked at 11% on Nov. 17, the day Apple and the NSA network failed.
On the main point, the experts matched what the operators later said: most failures traced back to the operators' own changes. Their AI theories moved with the news. First it was an AI routing tool the FAA had just rolled out, then AI-written code and staff cuts, then flaws found by AI that forced rushed patches. Only the last has a documented link, at the NSA. Everyone reached for a story within hours, officials included, and the officials' story arrived with more certainty than evidence. Gary Marcus, an AI scientist and frequent critic of the industry, told Fortune, "I doubt we are getting full candor", and added, "We don't really know yet."
If AI targeting can be tricked, why did the weapons talks stall?
Vals, a startup that evaluates AI models, tested leading commercial models on drone footage and satellite images. One test used real satellite images of a school in Minab, Iran, that the US struck on Feb. 28. Anthropic's Claude Opus 5.5 picked the school as a target in 69% of trials, and in 72% when told the task was a game. Telling models the task was a game was enough to get past refusals. Vals says it did not test how these models would perform inside a real weapons system.
The systems that steer machines can be fooled on purpose. A paper posted Nov. 3 and accepted to NeurIPS 2026, a major AI conference, tested Alpamayo, a self-driving model that explains its decisions in writing. Small changes spread across the camera image caused a collision in 52.6% of 154 test scenes with obstacles. In 89.5% of cases the model's written explanation stayed word for word the same, so reading what the system says it is doing would not reveal the attack. AI Weekly flagged the paper in Found First.
Lethal autonomous weapons are systems that can select and attack targets without a person deciding each strike. Governments have discussed limits on them for 13 years under a UN treaty, the Convention on Certain Conventional Weapons, which decides by consensus. When its review conference opened in Geneva on Nov. 16, Russia stopped the meeting before it started by objecting to observers, including the European Union, addressing it.
On Nov. 17 the talks moved into informal consultations. They are suspended, not over, and the conference runs through Nov. 20.
Russia blocked the procedure. The United States opposes the substance. The conference chair proposed two years of talks on a legally binding agreement; the US proposed five years of work toward a non-binding one. On Oct. 29 a UN General Assembly committee voted 163 to 5 to move toward negotiations; AFP reported that the five no votes were the United States, Russia, Belarus, Israel and North Korea. The US, with Russia, argues that the existing laws of war already cover these weapons.
What we still don't know, and who owes the answer
- Why both FAA sites failed at once. The FAA and its contractor L3Harris. No date given.
- The evidence for "complete certainty." Mullin, who testified to the House Homeland Security Committee on Nov. 18. We will report what he gave.
- Why Microsoft's tests missed a July defect. Microsoft's final report, due Nov. 20.
- How a signup flaw gave an attacker write access to 178 organizations' website code. Optimizely.
- Whether there will be weapons talks, and whether binding. The Geneva conference, by Nov. 20.
Key takeaways
- AI is not going rogue in the way people fear. People are deploying it faster than they can secure it. Attackers' AI agents still have a person in charge, and in targeting tests commercial models picked a school and were easy to trick.
- Washington ruled out an attack before it had a cause. The FAA still has no root cause, and Mullin's certainty rose without new public evidence.
- The outages were self-inflicted, and AI adds pressure. AI-found flaws force rushed patches; the NSA's network went down mid-patch.
- The one real attack came through a vendor and was found from outside. None of the 178 organizations has said it caught the code itself.
- AI is arming attackers with speed, not yet with independence. One China-linked group ran a free AI hacking tool against more than 200 organizations in 11 days, with a person in charge.
- The weapons talks stalled on two fronts. Russia blocked the procedure; the US opposes a binding treaty.
Worth Reading
- Cloudflare on AI hacking tools, with traffic numbers and what the tools look like in logs.
- Anthropic's red team on AI targeting, including the drone tests without GPS.
- The Alpamayo attack paper, on why a model's explanation cannot prove it is safe.
- Anthropic's new disclosure timeline, on why it is cutting fix windows to 30 days.
Wait, What?
- AI agents hacked countries they were told to avoid. In September the security firm GreyNoise tracked AI agents that broke into 395 organizations through flaws in PaperCut printing software, 11 of them within 26 seconds, including in countries the operator had told the agents to stay out of. The person running them set the goal. The agents did not stay inside the limits.
This week's poll
What worries you more about AI and security right now?
**What worries you more about AI and security right now?**
Reply and tell us what you saw at work during the outages.
AI Weekly special edition, compiled November 18, 2026 from operator statements, official comments, public reporting, research papers and AI Weekly's own Who's Who data. Written by Alexis.