Can AI Agents Hack Systems and Cause Outages? What's Real in 2026
Microsoft 365 went down on 12 November and part of AWS on 15 November, and both times people asked whether AI agents were behind it. Neither outage has been linked to AI. But AI agents are already part of real attacks, and their main goal is to steal passwords and accounts, not to knock services offline.
The short answer
- Have AI agents caused a major outage? No such case has been reported. Microsoft says its outage was a software defect, not AI or an attack. AWS has not said what caused its outage and has not described it as malicious.
- Are AI agents used in real attacks? Yes. Anthropic, Google and Mandiant have each documented cases where AI did much of the attack work.
- Can they run a whole attack on their own? Not yet, as far as anyone has confirmed. The International AI Safety Report (February 2026) says "fully autonomous end-to-end attacks, however, have not been reported," and Google said the same in September.
The outages that raised the question
Microsoft 365, 12–13 November. Outlook, Teams, SharePoint and Copilot were down for almost 24 hours. Microsoft's preliminary report blames a Windows defect from July that was triggered by extra traffic and reduced capacity, and it says there was "no indication" of malicious activity. Full story: Was the Microsoft 365 outage a cyberattack or AI?
AWS US-EAST-1, 15 November. From about 3:05 to 6:53 AM Pacific time, 11 services in AWS's Northern Virginia region had problems, including IAM Identity Center, NAT Gateway, Elastic Load Balancing and CloudWatch. Customers saw sign-in failures and connection errors. AWS has not published a root cause. Its status updates don't attribute the outage to malicious activity.
Two outages in four days at the biggest cloud providers makes people suspicious. But big cloud outages have happened regularly for years, long before AI agents, and the explanations published so far are the usual kind.
What AI agents have actually done in attacks
An espionage campaign run mostly by AI (2025)
In November 2025 Anthropic reported that a group it assessed with high confidence as Chinese state-sponsored had used Claude Code to try to break into about thirty organisations around the world. The AI did an estimated 80–90% of the work. Humans stepped in at perhaps 4–6 key decisions per campaign. It succeeded "in a small number of cases." The AI also made mistakes: it sometimes made up credentials, and claimed secrets that were actually public. Anthropic says that "remains an obstacle to fully autonomous cyberattacks."
Thousands of credentials stolen in under six hours (2026)
Google's Threat Intelligence Group (GTIG) described a case Mandiant investigated in the second quarter of 2026. An attacker who had broken into an organisation's cloud account set up a group of AI agents there. Using an AI coding chatbot, one prompt and a set of agent instructions, the attacker planned, built and ran a large credential-harvesting campaign in under six hours. It collected thousands of third-party credentials. Because it ran from the victim's cloud account, its traffic came from legitimate IP addresses.
Stolen secrets, collected automatically
GTIG also found an exposed server running a framework called "Recon" that held more than 23,800 stolen secrets, including cloud and AI service API keys. Google describes a shift from "passive, endpoint-focused infostealers to offensive agentic harvesting."
Why stealing matters more than breaking
Every confirmed case above follows the same pattern: get in, collect credentials, keys and data, and move on. That is where the money is. An outage gets noticed within minutes and draws a large response, while stolen sign-ins can be used or sold for weeks.
For most people and small businesses, the realistic risk from AI-assisted attackers isn't a cloud provider going dark. It's this:
- Better phishing, at volume. Well-written messages in your language that refer to real events, like the fake Microsoft outage notices sent during the 12 November outage.
- Fake "verify you are human" pages that tell you to press Win+R and paste a command. This is one of the most common ways password stealers get onto Windows PCs.
- Infostealers that copy the browser files holding your saved passwords and sign-in cookies. A stolen session cookie can let someone into your account without your password or your MFA code.
- Remote-access scams, where a fake support agent talks you into installing AnyDesk or TeamViewer.
How PhishClean Desktop helps on a Windows PC
Most of these attacks end with something happening on your PC, after you've left the web page, where a browser extension can't see. PhishClean Desktop watches for those steps:
- Fake CAPTCHA commands: when a page tricks you into running a command through Win+R, PhishClean ends it before it downloads anything. Free.
- Password and cookie thieves: any program other than your browser that opens the files where Chrome, Edge or Firefox store saved passwords and sign-in cookies is stopped on Pro. The free plan warns you.
- Cloud, code and AI keys: the kind of secrets found on the "Recon" server. A program run from Downloads, Temp or the Desktop that opens your AWS, Azure, Google Cloud, SSH, GitHub or npm keys, or your Claude, Codex, Gemini or Copilot sign-in, is stopped on Pro. Your own developer tools are left alone.
- Crypto wallets and chat sessions: the same guard covers MetaMask, Exodus, Electrum, Telegram and Discord.
- Screen-sharing scams: a full-screen warning if a remote-access app is running while you open your bank, UPI or a government site.
- Risky downloads: scripts, APK files and disguised programs are flagged as soon as they land in Downloads. Free.
What it doesn't do: PhishClean Desktop can't stop an attacker inside a cloud provider, and it doesn't replace your antivirus. Keep Microsoft Defender on. It doesn't detect AI agents as such. It looks for what an attack does on your PC, whether a person or an AI is running it.
A five-minute checklist
- Use a passkey or an authenticator app on your email account. SMS codes are the weakest option.
- Never paste a command into Win+R, PowerShell or Terminal because a web page told you to.
- Treat any message about a news event (an outage, a breach, a refund) as a possible lure, and check it through the company's own status page or app.
- Sign out of sessions you don't use, and remove browser extensions you don't recognise.
- Keep Windows and your browser up to date. AI is making it faster for attackers to exploit newly published bugs.
Sources
- Anthropic, 13 November 2025: the AI-orchestrated espionage campaign.
- Help Net Security, 8 September 2026: Google GTIG's AI threat tracker, the Mandiant six-hour case and "Recon".
- The Register, 3 February 2026: the International AI Safety Report on autonomous attacks.
- GBHackers, 16 November 2026: the AWS US-EAST-1 outage timeline and affected services.
- Office 365 for IT Pros, 16 November 2026: Microsoft's preliminary post-incident report.
Try PhishClean Desktop free for 7 days
Every protection on for 7 days on Windows 10 and 11, no card needed. Your browsing, files and clipboard stay on your PC.
Download PhishClean DesktopLast updated: