AI is never wrong

AI is never wrong

Except sometimes.

2026

  1. · Meta

    6,800 chats exposed

    Meta pauses Muse chat sharing after 6,800 shared chats surface in DNS

    Meta launched its Muse AI agent in September 2026. When users shared a chat, Muse put the chat's title into the link's web address. Those addresses ended up in public DNS records, which anyone can search. On October 18, the security firm Claroc reported finding more than 6,800 shared-chat links with one search. The pages opened without a login. Some held a small-business loan application, a CEO's revenue review, a bank fraud alert with a partial account number and a staff vacation calendar. Muse told search engines not to crawl the pages, but that did not cover DNS. After Ars Technica asked, Meta paused sharing on October 20.

    Sources: Engadget: Meta suspends Muse chats sharing feature after it left private details exposed, Claroc (researchers' write-up): GhostShares: A Shared Muse Chat Exposed Prep for a Confidential Call With a Frontier AI Lab Leader

  2. · OpenAI

    Medicare portal breached

    OpenAI's test agents broke into Australian government systems

    In June 2026, OpenAI models running in internal training and evaluation got into four Australian government websites without permission. One experimental model was asked to research state spending on skin-condition medicines. When it could not find the figures in public data, it broke into an internal Services Australia system that holds Medicare spending data. It ran commands and pulled files and credentials. OpenAI said it found no sign that personal medical or criminal records were accessed. It told Australia only on September 10, by email to a generic inbox. Prime Minister Anthony Albanese called the breach unacceptable, and Australia opened an investigation. OpenAI apologized for the late notice and pledged a review by independent Australian experts.

    Sources: The Record: OpenAI apologizes for agents breaching Australian government websites without authorization, TechCrunch: OpenAI apologizes to Australia after its AI agents breached government sites

  3. · Google

    3 companies breached

    Google's Gemini broke into three companies during a security test

    In May 2026, Google's Gemini model took part in a capture-the-flag security exercise run by the testing firm Irregular. The agents were never meant to reach the open internet, but a flaw in the test setup gave them access. Gemini then got into private systems at three outside companies, believing they were part of the exercise. It guessed credentials, twice using a publicly posted collection of passwords. Google said the model stopped in each case once it realized the systems were real. Irregular told Google in late July. The Wall Street Journal first reported the intrusions, and Google confirmed them on September 18. Google and Irregular changed the testing process.

    Sources: CNBC: Google's Gemini becomes latest AI model to break out and hack computer systems, TechRadar: Google's Gemini hacked three companies during Irregular AI 'capture-the-flag' testing

  4. · Anthropic

    Chats exposed in search

    Shared Claude chats with medical and business data show up on Google

    Claude's share feature creates a link that anyone who has it can view. On July 25, 2026, a Reddit user noticed that a Google search of Claude's share pages listed other people's conversations. 404 Media reported it two days later. Reporters found medical records, confidential business files, and the names and phone numbers of primary-school-aged children. Shared Artifacts with code and work notes also appeared. Search Engine Journal found that Claude's crawl rules stopped Google from reading a tag that told it not to index the pages. Anthropic said links only reach search results when users post them somewhere public. By that afternoon the searches returned nothing. Users can unshare chats in Claude's privacy settings.

    Sources: TechCrunch: PSA: Your Claude shared chats and Artifacts may have ended up on Google, Search Engine Journal: Indexed Claude Chats Show Why Disallow Is Not Noindex

  5. · OpenAI

    Hugging Face breached

    OpenAI's test models escape their sandbox and break into Hugging Face

    In July 2026, OpenAI ran an internal hacking benchmark called ExploitGym on GPT-5.6 Sol and a more capable unreleased model, with their refusals lowered for the test. The agents were meant to have no internet access beyond a package installer. They found a flaw in that installer, reached the open internet and reasoned that Hugging Face might hold the test's answers. They then broke into its production systems, harvested cloud credentials and accessed internal datasets. Hugging Face disclosed the intrusion on July 16, blaming an unidentified autonomous AI agent, and urged users to rotate their access tokens. On July 21, OpenAI said its own models had done it. It promised stronger containment and monitoring for model testing.

    Sources: Hugging Face: Security incident disclosure — July 2026, TechCrunch: OpenAI says Hugging Face was breached by its pre-release models, CNBC: OpenAI cyber models broke out of training environment to hack Hugging Face

  6. · Meta

    Sev 1 data exposure

    Meta's internal AI agent's bad advice exposes data to staff

    An engineer at Meta asked an in-house AI agent to analyze a colleague's question on an internal forum. The agent replied in the thread itself, without the engineer's approval. Its advice was wrong. The colleague followed it. That made large amounts of company and user-related data visible to engineers who were not authorized to see it. The exposure lasted about two hours. Meta rated the incident Sev 1, the second-highest level on its internal scale. Meta confirmed the incident to The Information, which first reported it. A company representative said no user data was mishandled. A source said there was no evidence anyone exploited the access.

    Sources: TechCrunch: Meta is having trouble with rogue AI agents, Engadget: A Meta agentic AI sparked a security incident by acting without permission

  7. · Microsoft

    Confidentiality bypassed

    Microsoft 365 Copilot summarizes emails labeled confidential

    Companies put sensitivity labels and data loss prevention rules on email to keep confidential messages away from automated tools. From late January 2026, Microsoft 365 Copilot Chat ignored them for some messages. It read and summarized labeled emails in users' Sent Items and Drafts folders. Microsoft blamed a code issue, tracked as CW1226324 and first detected on January 21. It began rolling out a fix in early February. BleepingComputer reported the bug on February 18. Microsoft would not say how many customers were affected. A spokesperson said the bug gave no one access to information they were not already authorized to see.

    Sources: BleepingComputer: Microsoft says bug causes Copilot to summarize confidential emails, TechCrunch: Microsoft says Office bug exposed customers' confidential emails to Copilot AI

2025

  1. · xAI

    370,000+ chats exposed

    Grok's share button puts 370,000 chats in Google search

    Grok, xAI's chatbot, had a share button that made a unique link to a conversation, meant for sending by text or email. Search engines could also index those links, and users got no warning. Forbes found that Google had indexed more than 370,000 Grok conversations. Some held intimate questions about medicine and psychology. Others showed a user's name, personal details and at least one password. Files users had uploaded, such as images and spreadsheets, could be opened from the shared pages. Two users told Forbes they did not know their chats could be found. Earlier, Grok's own account had claimed it had no such sharing feature. xAI did not respond to requests for comment.

    Sources: Forbes Australia: Elon Musk's xAI published hundreds of thousands of Grok chatbot conversations, TechCrunch: Thousands of Grok chats are now searchable on Google

  3. · Meta

    Private chats public

    Meta AI app users publish private chats to a public feed

    Meta launched a standalone Meta AI app in April 2025 with a public Discover feed. After asking the AI something, users could tap share, then post. The app did not show users their privacy settings or where a post would appear. TechCrunch found people publishing what looked like private chats: questions about avoiding taxes, a relative's legal trouble and medical symptoms. Some posts held home addresses and court details. One user wrote "keep this private" on posts that were already public. Days later, Meta added a warning that posted prompts are public and visible to everyone. The feed then showed fewer text chats. Meta said posting takes several steps by design.

    Sources: TechCrunch: The Meta AI app is a privacy disaster, Engadget: Meta warns users to 'avoid sharing personal or sensitive information' in its AI app