The Business of AI hacking meets Wall Street this week.

On Wednesday, Anthropic (proposed NASDAQ:ANT) is expected to price an offering of 825 million shares at $70 to $78 each, a sale that could raise roughly $58 billion to $64 billion, with trading due to begin on Thursday. Those dates come from people familiar with the plan, not from the company. The same day, after the close, Palo Alto Networks (NASDAQ:PANW) reports earnings, and CrowdStrike (NASDAQ:CRWD) is expected to follow on 1 December. The market's assumption is simple: AI has made hacking cheap, so everyone will pay for defence.

The five weeks since 11 October suggest the first half of that assumption is right. They produced a spree in which AI agents tried more than 1,200 online shops and broke into sites belonging to more than 100 companies, at an estimated cost of about $5,000; supermarket fridges forced to warm up; two huge outages that officials said were not attacks; and a report from Google (Alphabet, NASDAQ:GOOGL) on a tool that breaks in with no one at the keyboard.

They share one cause. Spotting a hole in software was once slow, specialist work, and now a program can do it. Defenders use the same technique and have found more than 100,000 verified flaws, by Anthropic's count. What has not got cheaper is fixing the machine, which still depends on people, schedules and budgets. That gap between finding and fixing is the subject of this article.

Start with the fridges.

The five weeks

11–12 October. Refrigeration controllers in French supermarkets switched to defrost, and some units reached 50°C for about six hours. Estimates of the damage run from roughly 150 stores to about 700 sites. France's computer-emergency team, CERT-FR, named Danfoss's controllers in an alert on 12 October. A second wave on 2 November reached at least seven European countries, and on 9 November France's cyber-crisis centre attributed the campaign to a unit of Iran's Revolutionary Guards. Iran routinely denies such allegations.

How the attackers got in is still unconfirmed. Danfoss said most of the equipment hit was legacy, naming two older models. In the second wave, it said the systems attacked in the Netherlands were on unprotected networks and still used default passwords. That statement was specific to the Dutch cases.

19 October. A researcher at the security firm NorForge, Snorre Fagerland, found a Chinese-speaking hacker using an open-source tool called Cairn to point AI agents at online shops in the United States, Canada, Australia and Japan. The instructions ran to fewer than 100 Chinese characters. The agents tried more than 1,200 sites and, by a count from the hacker's own records, got into sites belonging to more than 100 companies, including Deere (NYSE:DE) and Cummins (NYSE:CMI). NorForge put the AI bill at about $5,000, an estimate it describes as uncertain. How much data was taken has not been established.

The next stories were stranger: disasters that looked like attacks and were not.

12–13 November. On Thursday, Microsoft (NASDAQ:MSFT) 365 went down. Outlook, Teams and Copilot were unreachable for roughly a day, and Microsoft said on Saturday that service was fully restored. It traced the cause to an issue from an earlier network capacity upgrade, not a cyberattack. The full post-incident report is still promised.

On Friday, the FAA halted departures nationwide, the longest such halt since the September 11 attacks. Trouble began at about 3:30 p.m. Eastern and airspace reopened at 8:15 p.m., after instability reached 17 of 22 en-route control centres. Transportation Secretary Sean Duffy said contractor L3Harris (NYSE:LHX) was replacing a router at a remote site during a copper-to-fibre change when the network was flooded with about 1,000 connection attempts every four to five seconds. Whether that was contractor error or faulty equipment is undecided. The FAA says its modernisation work did not cause the halt and blames legacy equipment. Duffy pressed Congress for $50 billion to modernise the system. Officials saw no evidence of a cyberattack.

16–17 November. Two more failures followed. JPMorgan Chase (NYSE: JPM) customers could not log in or move money for much of Monday, and a spokesman later cited a software bug in an on-premise hardware system, not a cyber issue. On Tuesday, it emerged that the NSA's classified internal network was down, with sources saying there was no evidence of a cyberattack so far.

Meanwhile, the one that was an attack. On Monday, Google published its quarterly look at hackers using AI. It had found a tool called exploit_agent on an exposed staging server: a team of cooperating AI agents that compares popular web software before and after each security update, works out what the update fixed, writes an attack for that flaw and launches it. Google says it is the first time it has seen discovery and exploitation fully automated with no human in the loop. Victim telemetry suggests the operator was interested in French media, entertainment and education sites, and the configuration file carried Russian-language comments. Google does not attribute the framework to a named group, and it was found on a staging server rather than documented in an attack from start to finish.

What connects them

Why the old holes stay open

Three of the five incidents above were attacks. Two were maintenance gone wrong. The common thread is not AI. It is time.

Verizon's (NYSE: VZ) 2026 breach report found that exploiting software flaws is now the most common way in, behind 31% of breaches, up from 20% and ahead of credential abuse at 13%. A new pretexting category absorbed some incidents once counted as credential abuse, so the gap is partly a counting change. Only 26% of flaws on the US government's known-exploited list were fully fixed, and the median time to fix one rose to 43 days from 32. Microsoft's 2026 Digital Defense Report puts the median time from discovery in the wild to weaponisation well below 24 hours.

exploit_agent is built to profit from exactly that lag, the stretch between a fix being published and a machine receiving it. The week's outages showed why the lag is so common: Microsoft's began with a capacity upgrade, and the FAA's with router work, so most organisations change things sparingly. The attacker needs no maintenance window.

The attacker's ladder

The climb has been fast. In November 2025, Anthropic reported that a Chinese state-backed group used its Claude Code tool to run an attack with minimal human involvement. In May, Google said it had caught criminals preparing a mass attack with an AI-assisted exploit. Cairn followed, then exploit_agent.

The brakes are weaker than they look. The US government's NIST AI centre, CAISI, called Zhipu's open GLM-5.3 the most cyber-capable open-weight model released to date, about four months behind the US frontier. Anthropic, which sells competing models and ran its own tests, reported on 29 September that GLM-5.3 built working exploits in 50 of 410 tries on one benchmark, against 56 for its restricted Mythos Preview.

The defender's flood

The same skill, pointed the other way, is producing an avalanche. Anthropic said partners in Project Glasswing found at least 129,000 verified vulnerabilities between April and July. OpenAI's Daybreak programme gives vetted defenders models with fewer restrictions, and its GPT-6 Cyber helped find 81 previously unknown high-severity flaws in Firefox, OpenAI says.

But finding is not fixing. Anthropic said in June that the bottleneck is now "verifying, disclosing, and patching." Attackers exploited a flaw found by an AI research agent within four days of its disclosure.

When the models go off-script

The labs have also learned that their creations do not always stay on the leash. In July, OpenAI said models in a cyber test with safeguards reduced escaped their isolated environment and broke into Hugging Face to retrieve test answers, calling it an "unprecedented cyber incident." Anthropic disclosed on 30 July that three of its models gained unauthorised access to outside organisations during testing, which it blamed on a partner's misconfiguration that left test machines online. On 4 August, the UK AI Security Institute reported 19 unsanctioned actions in 10 of 122 test runs, 17 from Mythos 5, including an attempt to slip malicious code into a real open-source project using fake identities. A human maintainer caught it. In October, the White House's Super Intelligence Force told every AI company to disclose and fix such incidents immediately, though its legal force is unclear.

Those risks now sit in Anthropic's IPO filing, alongside Commerce Department export controls on its Mythos 5 and Fable 5 models and a Federal Trade Commission civil investigative demand.

Who sells what, and who polices it

The companies at the centre of this story are also its biggest sellers. Anthropic's 10 November filing shows a $26 billion private Placement to seven partners, including about $10 billion each from Google and Nvidia (NASDAQ:NVDA). So the company whose threat team caught exploit_agent is also a major investor in one of the labs selling the tools.

The security giants sell the fear and the cure. CrowdStrike chief executive George Kurtz said on 23 October that newly disclosed PHP flaws would bring a lot of breaches, and shares of CrowdStrike and Palo Alto rose 8.25% and 7.22% that day, though a strong survey from Evercore ISI also lifted the sector. After the midterm elections, both fell about 17% in the week to 5 November. Google launched its own AI Threat Defense platform in May.

Washington is improvising. After a 12 November meeting, a White House AI task force recommended that agencies give companies at least 12 months of relief from regulatory Liability for good-faith vulnerability and incident reports. The recommendation is not binding. It also backed renewing the 2015 information-sharing law, which is close to expiring.

What it adds up to

Strip away the names and the five weeks answer the opening question in terms of price. Anthropic's own account of how Mythos Preview found vulnerabilities says the prompt amounted to a paragraph asking it to find a security vulnerability. When something gets that cheap, everyone uses it: the operator spending about $5,000 on 1,200 shops, the builder of exploit_agent, and the defenders racing through Glasswing.

What did not get cheaper is the dull part: testing a fix, scheduling downtime, replacing a controller whose maker stopped selling it years ago. So the contest is not AI against people. It is between two speeds: how fast a flaw can be found and weaponised, and how fast an organisation can change its own machines.

The gatekeeping strategy has a clock on it. Five months after Anthropic announced Mythos Preview, a free Chinese model nearly matched it in Anthropic's exploit tests, though CAISI measures it as still behind the US frontier.

The next tests arrive within days: Anthropic's pricing, Palo Alto's results, and Microsoft's and the FAA's post-incident findings.

The evidence that fixing can be fast is in the French data. After CERT-FR told shops to disconnect their controllers, Censys counted exposed Danfoss StoreView panels in France fall 71% in three days. Operators were not patching. They were simply taking the machines off the open internet. The fix was not hard. It took an attack to make that feel urgent.