Between October 19 and 22, 2026, security teams faced a wave of disclosures that people online quickly called "Black October". Among the hacked servers and poisoned software updates were two stories every business with a chatbot should read: private AI conversations found on the open internet, and stolen keys to AI accounts.
What happened
In a few days, a critical flaw in PHP (nicknamed PHP2Shell) led sites such as the Internet Archive to go offline as a precaution. Two cloud hosting platforms, Railway and Zeabur, said attackers may have read passwords and keys their customers had stored. A joint warning from US and European agencies said one ransomware gang had hit more than 600 victims since 2024. Many of these attacks had started weeks or months earlier and only came to light now, which is part of why it felt like everything happened at once.
Some of the claims that spread online were exaggerated. The incidents themselves are real.
Your customers' chats are data
On October 20, researchers reported that shared conversations from Meta's AI assistant, Muse, could be found through public internet records. One search turned up 6,800 links. Among what people had shared: sales reports, a small-business loan application and a bank fraud alert with partial account details. Meta suspended sharing while it investigates.
Those users weren't hacked. They typed sensitive information into a chatbot, shared the conversation, and it ended up somewhere they didn't expect. Your website's chatbot works the same way: customers type names, addresses, symptoms, card numbers and complaints into it, and every one of those conversations is stored somewhere.
The keys behind the chatbot are a target
On October 21, attackers used stolen credentials to publish a version of LiteLLM, a popular tool that connects applications to AI models, that stole passwords, cloud logins and AI account keys. It was online for less than two hours on one download site and about two and a half on another. Many chatbots are built on tools like this one.
A stolen key lets someone use your AI account in your name and run up your bill. In a separate case, criminal groups used an open-source AI agent tool to target more than 1,100 online shops and companies, and broke into more than 100.
Most chatbots have never been tested for this
Most chatbots are checked to see that they work, not to see what they give away. Courts and tribunals already treat a chatbot's words as the business's own: in 2024 a tribunal made Air Canada honor a refund its chatbot invented. If yours can be talked into sharing a customer's details, that is your leak, not your vendor's.
Sources
- ScruteX: Black October 2026, 28 cyber incidents checked (Oct 23, 2026)
- WebThreat: Black October 2026, what leaked
- GIGAZINE: Over 6,800 Meta Muse shared links found (Oct 21, 2026)
- SecurityToday: Black October, many attacks had been running for months (Oct 22, 2026)
- The State of AI: Inside Black October, and the AI behind it
- PCWorld: Meta just reminded us that sharing AI chats is a terrible idea
- IBTimes: Researchers uncover 6,800 Muse share links