AI, data breaches and your private messages
Facts checked November 10, 2026. This page reports published incidents and studies. It predicts no breach of any service, and it is not legal or security advice.
EverPosted's free scan runs in your own browser and shows which of your old posts touch a sensitive topic. You download your own data export, and the scan uploads nothing from your files.
The short answer
Researchers using AI models now find serious flaws in old, widely used software. Breaches are counted in the thousands every year, and criminals have used AI models to sort what they stole. Each of those statements has a source below.
What follows from them is EverPosted's own reasoning: know what each service holds about you, and delete what you no longer need.
AI is finding old flaws in common software
- November 2024. Google's Big Sleep agent found an exploitable flaw in SQLite before it was released. Google wrote: “We believe this is the first public example of an AI agent finding a previously unknown exploitable memory-safety issue in widely used real-world software.”1
- August 2025. In DARPA's AI Cyber Challenge final, “teams also discovered 18 real, non-synthetic vulnerabilities that are being responsibly disclosed to open source project maintainers.”2
- March 2026. A FreeBSD security advisory credits “Nicholas Carlini using Claude, Anthropic.”3 Anthropic describes the flaw as “a 17-year-old remote code execution vulnerability in FreeBSD that allows anyone to gain root on a machine running NFS.”4
Most flaws that are found never get used. Google's threat researchers report that “only 0.23% of all disclosed vulnerabilities in 2026 (roughly 1 in 431) were ever observed in active exploitation.”5
October 2026: the PHP flaws
On October 21, 2026, PHP released versions that “fix two critical vulnerabilities that can allow unauthenticated remote code execution.”6 CISA's alert says the pair is known as “php2shell.”7 PHP runs the server side of 68.9% of the websites whose programming language W3Techs can identify.8
- Who found them. The CVE records credit Calif, Wiz and OpenAI.9 OpenAI says Calif found one flaw with an OpenAI model, “with Wiz also independently discovering and reporting both vulnerabilities.”10 Wiz says it found the first “using Wiz Atlas, Gemini 4 Argon, and GPT-5.6 Sol.”11 The PHP Foundation says a PHP core developer had already found “the first signs” of one problem, and that in 2024 “a potential exploit for the OOM problem was theorized, but no concrete vectors were found.”12
- How old. Calif wrote: “The earliest affected path we found dates to PHP 5.4.0.”13 PHP 5.4.0 was released on March 1, 2012.14
- What got out early. The day before the release, “the two CVE IDs began circulating on online forums and social media.” The Foundation says “No technical details leaked,” and it moved the release up by a day.12
- Attacks. Amazon “identified a PHP2Shell exploitation attempt dating to Oct. 17, 2026,” four days before the flaws were public.15 CISA added both flaws to its catalog of vulnerabilities known to be exploited on October 26.16 None of these sources counts how many websites were broken into.
OpenAI wrote: “To our knowledge, frontier AI found these vulnerabilities before attackers did.” It also wrote that “the same capabilities can also help attackers rapidly develop offensive tools and exploits,” and that it expects “similar findings to increase sharply in the coming months.”10
Breaches are routine
The Identity Theft Resource Center counted 3,322 data compromises in the United States in 2025, the most it has recorded in a year, and 5% more than in 2024. Notices to victims fell to 278.8 million, because 2025 had none of the “mega-breaches” of 2024. Seven in ten breach notices did not say how the breach happened. In the Center's own survey of 1,040 consumers, 80% had received a breach notice in the last 12 months.17
Verizon's 2026 report covers “more than 22,000” confirmed breaches in 145 countries. It says “Exploitation of vulnerabilities is now the most common initial access vector for breaches. It has risen to 31% in this year's reporting dataset.”18
What has been exposed besides passwords
- Searches and places, Facebook, 2018. Attackers stole access tokens for about 30 million people. For 14 million of them they reached profile details that included “the last 10 places they checked into or were tagged in” and “the 15 most recent searches.” For about 400,000 people they also saw “the names of recent Messenger conversations.” Facebook said “Message content was not available to the attackers, with one exception.” The exception was messages received by a Page whose admin was in that group.19
- Direct messages and data downloads, Twitter, 2020. After a phone phishing attack on employees, “the attackers targeted 130 Twitter accounts, ultimately Tweeting from 45, accessing the DM inbox of 36, and downloading the Twitter Data of 7.”20 The last of those is the account's own data export.
- Chat titles, ChatGPT, 2023. This one was a bug, with no attacker involved. OpenAI said it “allowed some users to see titles from another active user's chat history,” and that the first message of a new conversation may also have been visible to someone else.21
- Private communications, US telecoms, 2024. The FBI and CISA described “the theft of customer call records data, the compromise of private communications of a limited number of individuals who are primarily involved in government or political activity.”22
- Archived chats, TeleMessage, 2025. The CVE record says the archiving backend “holds cleartext copies of messages” from users of its TM SGNL app, “as exploited in the wild in May 2025.”23
- Support conversations and IDs, Discord, 2025. An attacker got into a customer service vendor's systems “with a view to extort a financial ransom from Discord.” The data included “Messages with our customer service agents,” and “approximately 70,000 users that may have had government-ID photos exposed.” Discord said “No messages or activities were accessed beyond what users may have discussed with Customer Support or Trust & Safety agents.”24
Stolen data gets sorted with AI
- August 2025. Anthropic reported a criminal who “targeted at least 17 distinct organizations, including in healthcare, the emergency services, and government and religious institutions.” Anthropic's model “was allowed to make both tactical and strategic decisions, such as deciding which data to exfiltrate, and how to craft psychologically targeted extortion demands.”25
- November 2025. Anthropic reported a campaign against “roughly thirty global targets” that “succeeded in a small number of cases.” The attackers used the model to “extract a large amount of private data, which it categorized according to its intelligence value.”26
- September 2026. Google reported a data theft and extortion group using an AI coding tool “to parse and analyze exfiltrated directories for extortion.”27
These are a handful of documented cases. In the same report Google wrote that it “has not yet observed threat actors deploying fully autonomous pipelines against targets in the wild.”27
What you can do
After a breach notice, the FTC's steps are about credit and logins. Check your credit reports, and consider a freeze: “A credit freeze keeps people from getting into your report.” The FTC adds that “it's free to place the freeze, or to temporarily lift the freeze if you need to apply for new credit.” Change the password, and “If you use the same password anywhere else, change that, too.”28
Those steps do not change what a service already stores about you. The rest of this list is EverPosted's reasoning, with an agency's words where one has spoken.
- See what each service holds. Each of the six platforms EverPosted covers will send you a copy. What's in your data export says what comes back.
- Close accounts you no longer use. NIST's advice to individuals: “Close old accounts. If you don't use them, close them.” NIST marks that page as no longer updated.29
- Delete posts and messages you don't need, after you download a copy. What a service no longer stores cannot be taken from it later. How to delete old posts covers what each platform lets you remove, and when to ask a lawyer first.
- Treat public posts as readable by anyone. CISA says “Threat actors can use your publicly available information to launch targeted cyber intrusions.”30
Where to start
- Request a copy of each account. How long each takes:
- Run the free scan as each one arrives. It runs in your own browser and uploads nothing from your files. It counts the posts and messages that touch violence, drugs, sex, legal trouble, harsh words, and your phone numbers and addresses.
- Decide what to delete. The EverPosted report reads each flagged item in context and ranks them. The background check page covers what a screener looks for today.
Sources
- Google Project Zero, “From Naptime to Big Sleep: Using Large Language Models To Catch Vulnerabilities In Real-World Code”, November 1, 2024. projectzero.google
- DARPA, “AI Cyber Challenge marks pivotal inflection point for cyber defense”, August 8, 2025. darpa.mil
- The FreeBSD Project, Security Advisory FreeBSD-SA-26:08.rpcsec_gss, March 26, 2026. freebsd.org
- Anthropic, “Claude Mythos Preview's cybersecurity capabilities”, April 7, 2026. anthropic.com
- Google Threat Intelligence Group, “Vulnerability Discovery and Exploitation Trends in the AI Era”, September 30, 2026. cloud.google.com
- PHP, news archive for 2026, security release of PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35, October 21, 2026. php.net
- CISA, “Critical Vulnerabilities in PHP Require Immediate Patching”, October 21, 2026, revised October 26, 2026. cisa.gov
- W3Techs, “Usage Statistics and Market Share of PHP for Websites, November 2026”, November 10, 2026. w3techs.com
- CVE Program, records CVE-2026-109618 and CVE-2026-109619, October 21, 2026. cve.org
- OpenAI, “Patch the Planet: global response coordination as agents uncover critical vulnerabilities”, October 21, 2026. openai.com
- Wiz Research, “PHP2Shell: AI uncovers critical RCE vulnerabilities affecting 70% of the Internet”, October 21, 2026. wiz.io
- The PHP Foundation, “Today's PHP Security Release, and How We're Strengthening PHP Security”, October 21, 2026. thephp.foundation
- Calif, “PHP2Shell by Calif Research”, October 21, 2026. calif.io
- PHP, PHP 5 ChangeLog, version 5.4.0, March 1, 2012. php.net
- Amazon Web Services, “Amazon Threat Intelligence identifies PHP2Shell exploitation prior to public disclosure in MadPot”, October 21, 2026, updated October 22, 2026. aws.amazon.com
- CISA, Known Exploited Vulnerabilities Catalog, catalog of November 9, 2026. cisa.gov
- Identity Theft Resource Center, “Identity Theft Resource Center 2025 Annual Data Breach Report: Record Number of Data Compromises in 2025; 79 Percent Jump Over Five Years”, January 29, 2026. idtheftcenter.org
- Verizon, 2026 Data Breach Investigations Report, Executive Summary, 2026. verizon.com
- Meta, “An Update on the Security Issue”, October 12, 2018. about.fb.com
- X (then Twitter), “An update on our security incident”, July 30, 2020. blog.x.com
- OpenAI, “March 20 ChatGPT outage: Here's what happened”, March 24, 2023. openai.com
- FBI and CISA, “Joint Statement from FBI and CISA on the People's Republic of China (PRC) Targeting of Commercial Telecommunications Infrastructure”, November 13, 2024. cisa.gov
- CVE Program, record CVE-2025-47729, May 8, 2025. cve.org
- Discord, “Update on a Security Incident Involving Third-Party Customer Service”, October 3, 2025, updated October 9, 2025. discord.com
- Anthropic, “Detecting and countering misuse of AI: August 2025”, August 27, 2025. anthropic.com
- Anthropic, “Disrupting an AI-orchestrated cyber espionage campaign”, November 13, 2025. anthropic.com
- Google Threat Intelligence Group, “GTIG AI Threat Tracker: From Prompting to Autonomy”, September 8, 2026. cloud.google.com
- Federal Trade Commission, IdentityTheft.gov, “What To Do if Your Information Was Lost or Stolen, or Part of a Data Breach”, undated. identitytheft.gov
- NIST, “Digital identity for individuals”, updated June 3, 2026. nist.gov
- CISA, “Manage your Online Presence”, undated. cisa.gov