
For most of the history of cybersecurity, the attackers who truly kept CISOs awake were a small club. Chaining a handful of vulnerabilities into a working intrusion, moving laterally without tripping alarms, and getting out with the data took years of craft. That talent lived mainly inside a few state intelligence units and a few top-tier criminal crews. Everyone else, the script kiddies, the opportunists, the commodity ransomware affiliates, could be stopped by reasonable hygiene.
That assumption broke in 2026. The skill is no longer the bottleneck. AI-enabled attack agents now perform much of the tradecraft that used to require an elite operator, which means the number of people who can act like elite operators has exploded. The attacks themselves look familiar. What changed is how many people can run them, and how fast.
What the CEO of Palo Alto Networks is seeing
Nobody has put it more plainly than Nikesh Arora, chairman and CEO of Palo Alto Networks, whose incident response teams sit on the front line of thousands of breaches. On 2 November he posted that “Attacks will accelerate from here on,” and then made the point that frames this whole post:
“Attacks are not becoming more sophisticated, it’s the same attacks, faster and by a lot more people.”
Nikesh Arora, CEO, Palo Alto Networks, on X, 2 November 2026
In the same post, Arora recalled predicting six months earlier that Mythos-class capability would become widely available, and said it is now happening. Attack-path discovery that once took trial and error has become trivial with open-source LLMs, more people are getting good at hacking because it has become so easy, and you are only as secure as your supply chain.
This is not a CEO guessing. In April 2026 Arora ran Anthropic’s then-unreleased Mythos model against Palo Alto’s own internal systems, and it correctly flagged 70% of the vulnerabilities. When a model can find most of the flaws in a security company’s own estate, the question stops being whether attackers will use the same capability and becomes how many of them will.
Fourteen months of evidence
The shift from “AI helps hackers write phishing emails” to “AI runs the intrusion” happened fast, and it is well documented.

September 2025, GTG-1002. Anthropic disclosed that a Chinese state-sponsored group had turned Claude Code into an autonomous attack orchestrator that executed roughly 80 to 90% of the tactical work, with humans only approving key steps. About 30 organizations were targeted and a handful breached.
July 2026, Hugging Face. The largest hub for open AI models disclosed a production intrusion driven end to end by an autonomous agent, which entered through a malicious dataset and logged more than 17,000 actions over a weekend. OpenAI later said the agent was its own models escaping a sandboxed evaluation. Hugging Face reconstructed those events with AI in about an hour, work that would normally take days.
October 2026, Storm-0940 and ARTEX. Microsoft reported that a China-aligned actor used ARTEX, an open-source agentic penetration-testing framework that can drive any AI model, to automate attacks on hundreds of organizations. After compromising ConnectWise’s Cloud Backup application, the actor attempted sign-ins to more than 2,250 Microsoft 365 tenants at 2,900 attempts per minute, and several attacks reached privileged cloud access in under three hours. Within weeks, Microsoft saw a Cambodia-based, financially motivated group, Storm-3358, aiming the same framework at financial firms.
That last detail is the democratization story in miniature. A tool built for defenders, released as open source, moved from a state actor to a criminal crew inside a single quarter.
Just as important is what did not need AI. The roughly $6.3 billion Binance theft in late October came through a compromised DevOps engineer’s laptop, and a malicious LiteLLM package sat on PyPI for about two hours carrying a credential stealer. Identity, supply chain and edge devices remain the doors. AI just means far more people are trying the handles, far faster.
The last ten days: a resilience stress test
Between 10 and 17 November, an unusual run of outages hit services much of the economy depends on: Apple iCloud, Microsoft 365 for more than a day across 70+ countries, a nationwide FAA ground stop, Workday payroll, Chase and other banks, United Airlines, and Apple services again on 17 November.
To be clear, every one of these companies attributed its outage to a technical cause, and none classified it as a security incident. Officials also described the events as unrelated. Running services at this scale is genuinely hard, and these teams restored service quickly. I mention the week not to second-guess anyone, but because it is the most vivid recent lesson in how dependent we have all become on a small number of platforms.

Check Point drew the right conclusion: availability is part of security, and concentration risk is security risk, because a failure at one major provider stops organizations that had nothing to do with it. Moments of disruption also create openings. After the 2024 CrowdStrike incident, CISA warned of phishing and fake “fixes”, because confusion and urgency make social engineering easier.
Meanwhile, attackers did not take the week off. Vanderbilt University, Apify, ConnectWise ScreenConnect partners and Uber (via a supplier) all disclosed incidents. On 16 November, Optimizely found an attacker had exploited a vulnerability in its Experimentation product to inject malicious JavaScript into some customer projects; Report URI found infected scripts across 178 organizations, built to skim cards and steal cookies. Optimizely patched within about three hours.
The same day, Google’s Threat Intelligence Group reported exploit_agent, a framework that compares code before and after a security patch, writes the exploit and launches it with no human involved, the first such case Google says it has observed. Put the two halves of the week together and you have a preview of 2027: platforms we all depend on can stop for purely technical reasons, while machine-speed attackers get better at turning every new patch into a weapon.
The numbers behind the shift
The large threat-intelligence datasets of 2026 all point the same way:
- CrowdStrike 2026 Global Threat Report: activity from AI-enabled adversaries rose 89% year over year. Average eCrime breakout time, from initial access to lateral movement, fell to 29 minutes, and the fastest on record was 27 seconds. In one intrusion, exfiltration began four minutes after initial access.
- Mandiant M-Trends 2026: the estimated mean time to exploit a vulnerability is now minus seven days, meaning exploitation routinely starts before a patch exists. The hand-off from an initial-access broker to the next threat group fell from more than eight hours in 2022 to 22 seconds in 2025.
- IBM Cost of a Data Breach 2026: one in four malicious breaches now involves AI, a 56% increase over the previous year, and those breaches cost about $6 million on average, against a global average of $4.99 million.

A defender with a weekly scan and a monthly patch window is now planning around a timeline the attacker abandoned years ago.
A note of balance: “AI cyberattack” is not yet a standardized reporting category, human-led, AI-assisted activity is still more common than fully autonomous operations, and Anthropic noted that the AI in GTG-1002 hallucinated. The machines do not need to be flawless, only cheap, tireless and good enough at scale.
What the next 6 to 12 months will bring
The most important forecast came from the AI labs themselves. On 27 August, OpenAI, Anthropic, Google, Microsoft and more than 100 other organizations signed an open letter warning that, in the coming months, AI-enabled cyberattacks will become far more widespread and sophisticated as models around the world grow more capable, and that defenders have a limited window to act. When the companies building the models say this in public, believe them. Combining their warning with the forecasts from Google Cloud, Palo Alto Networks and others, here is what I expect through late 2027:
- Volume on known weaknesses. Expect more attacks, not cleverer ones. Every unpatched edge device, exposed secret and over-permissioned app identity will be found and tried, repeatedly, by many more actors.
- The patch window disappears. Google’s exploit_agent finding shows patches themselves becoming the attacker’s roadmap. Google Cloud’s 2027 forecast warns that multi-stage attacks may unfold in seconds, making weekly scans and monthly patch cycles ineffective.
- Supply chain and AI tooling become the favorite targets. Google expects attackers to poison open-source packages, MCP tools and libraries that AI coding assistants pull into codebases. The LiteLLM, ConnectWise and ScreenConnect incidents are the preview.
- Agent sprawl and decision manipulation. Enterprises are deploying agents faster than they govern them, creating a new form of shadow IT. Google also flags a shift from stealing data to quietly poisoning the data or logic that autonomous agents rely on, so they make harmful decisions at scale.
- Security debt comes due. Arora estimates about $1 trillion of global cybersecurity debt, and argues that nothing deployed seven or ten years ago is ready to handle AI. Legacy perimeter gear and manual SOC processes will be where the big breaches of 2027 start.
My own expectation, as someone who spends his days in infrastructure: the defining incidents of the next year will not be exotic zero-days. They will be agentic frameworks abusing legitimate credentials and trusted third-party access at machine speed, followed by attacks on backups and recovery systems to remove the victim’s ability to say no.
How to prepare: a practical enterprise playbook
Arora’s own prescription is refreshingly unglamorous: test your infrastructure for easily exploited vulnerabilities and misconfigurations, fix your own code and require the same of your vendors, update old or insecure perimeter infrastructure first, and put a security framework around new AI tools such as coding assistants. Here is how I would turn that into a 12-month plan.

In the next 90 days: shrink what can be found
- Attack yourself first, with AI. Run AI-assisted discovery against your external and internal estate before someone else does. If a model found 70% of a security vendor’s flaws, it will find yours.
- Retire end-of-life edge devices. VPNs, firewalls and gateways remain the first door tried. Replace what can’t be patched quickly.
- Purge long-lived secrets. Microsoft has observed agentic frameworks automatically hunting exposed credentials in public GitHub repositories since August. Scan, rotate and shorten token lifetimes.
- Raise your guard during any outage. Expect impersonation phishing and fake fixes when a major provider is down. Have the help desk verify callers and scrutinize password and MFA resets.
- Audit OAuth and app-identity grants. Storm-0940 rode a trusted third-party application identity into thousands of tenants. Know which apps hold which permissions in your cloud tenants.
In 3 to 6 months: lock down the supply chain and your agents
- Review every vendor with access. Inventory third parties with standing access, cut what is not needed and require security commitments from those that remain.
- Pin and vet open-source dependencies. A two-hour malicious package window is enough when pipelines pull automatically.
- Treat every AI agent as an identity. Google recommends giving each agent verifiable credentials and least-privilege access, and logging its inputs, reasoning, tool calls and policy compliance so decisions can be audited.
- Map dependencies and rehearse degraded mode. Know what breaks if your identity, SaaS, cloud or payment provider is down for 24 hours, run the tabletop, and set up out-of-band communications.
- Red-team agents before deployment. Make adversarial testing of agentic systems a mandatory step, not an afterthought.
In 6 to 12 months: defend and recover at machine speed
- Automate detection and containment. With breakout times measured in minutes, isolation of a host or revocation of a token has to happen without waiting for a human ticket. IBM’s data consistently shows lower breach costs for organizations that use AI and automation in security operations.
- Put AI agents in the SOC. Hugging Face’s hour-long reconstruction of 17,000 attacker actions is the model: AI to fight AI, with humans supervising.
- Protect the ability to recover. Keep immutable, isolated copies of critical data, and rehearse full restores, not just backup jobs. Attackers already target recovery infrastructure; make sure your last line of defense is not reachable with the same credentials as your first.
- Change what the board measures. Report time to patch actively exploited vulnerabilities, the share of edge infrastructure that is end of life, mean time to contain and number of ungoverned agents. Speed is now the metric that matters.
The bottom line
The uncomfortable truth of 2026 is that the elite tier of attackers did not get smarter. It got bigger. CrowdStrike, Mandiant, IBM, Microsoft, Google and the AI labs all point the same way: more attacks, at machine speed, through doors we already know about. The last ten days add one more lesson: resilience is part of security.
That is actually good news for prepared organizations. The doors are known. The fixes are known. What has changed is the deadline. The defenders’ window the AI industry warned about in August is open now, and it will not stay open for long.
Related on Lots of Data: Find, Fix, Repeat (Patch Calendar) · VM Escape and AI Agent Sandboxes · Agent Sandboxes and Storage
