AI Doxing: How Artificial Intelligence Turns Posts and Photos Into Names and Home Addresses — and How to Protect Yourself

AI has cut the cost, time and skill needed to expose people online; an analysis of how AI doxing works, what research shows, who is targeted, the law, and practical steps for individuals and organisations.
Executive Summary
Doxing, the practice of exposing someone's identity, address or other private details without consent, used to require time, patience and investigative skill. Artificial intelligence has removed most of those barriers. Tasks that once took an experienced investigator days can now be completed in minutes by someone with no training, a smartphone and access to widely available AI models.
Our key findings:
- The cost of de-anonymisation has collapsed. Peer-reviewed research shows large language models can infer personal attributes such as location, income and gender from ordinary online posts with up to 85% accuracy, at roughly one hundredth of the cost and 1/240th of the time a human analyst would need.
- Photos have become a major leak. Multimodal AI models can now estimate where a photo was taken more accurately than non-expert humans, including private photos such as selfies taken at home.
- Face search closes the loop. Combining facial recognition search engines with AI-driven data aggregation can turn a single image of a stranger into a name, a phone number and a home address.
- The threat is broad. Journalists, activists, executives, public officials, security researchers, women in public life and ordinary social media users are all realistic targets.
- Defence is possible, but it must be deliberate. Reducing what you publish, removing data from people-search sites, cleaning photos and separating identities significantly raise the effort required to dox you.
What Doxing Is, and What AI Changed
Doxing is not hacking in the traditional sense. Most doxing relies on information that is already public or semi-public: social media posts, photos, old forum accounts, company websites, property and business registries, data broker listings and leaked databases. The harm comes from aggregation. Individually harmless facts become dangerous when they are combined into a profile that links an online identity to a real person and a physical location.
Historically, three constraints limited doxing:
- Skill. Linking a pseudonymous account to a real person required investigative experience and familiarity with open-source intelligence (OSINT) techniques.
- Time. Reading years of posts, cross-referencing usernames and checking photos for clues took hours or days per target.
- Scale. Because each target required manual effort, attackers could pursue only a few people at a time.
AI weakens all three. Language models read and reason over large volumes of text instantly. Vision models analyse photos for location clues. AI agents can run searches, follow links and assemble findings without human supervision. The result is not a new kind of attack but a dramatic change in its economics: doxing becomes cheaper, faster, more accurate and available to far more people.
How AI Doxing Works
AI-assisted doxing typically follows a chain of steps. Each one is now partly or fully automated.
- Collection. The attacker gathers everything linked to a target handle, email address, photo or name: posts, comments, profile pictures, usernames on other platforms and leaked records.
- Attribute inference. A language model reads the collected text and infers details the person never explicitly shared, such as their city, age range, occupation, relationship status or employer, from writing style, local references, time zones and casual remarks.
- Image geolocation. A vision model examines photos for landmarks, vegetation, road markings, architecture, signage, weather and sun angle to estimate where they were taken, sometimes down to a neighbourhood or street.
- Face matching. A facial recognition search engine matches a face against images across the web, revealing other profiles, news articles or event photos that carry the person's real name.
- Enrichment. The real name is then fed into people-search sites, data broker listings, business registries and breach data to find phone numbers, home addresses, relatives and employers.
- Publication or use. The final profile is posted publicly to intimidate the target, or used privately for harassment, stalking, extortion, fraud or physical threats.
Evidence From Research
Several studies and demonstrations show how capable these tools have become.
- Inference from text. A study by researchers at ETH Zurich, presented at the ICLR 2024 conference, tested nine widely used language models on real Reddit profiles. The best models inferred personal attributes such as location, income and sex with up to 85% top-1 and about 95% top-3 accuracy, at around 100 times lower cost and 240 times less time than human annotators. The authors warned that standard anonymisation of text does little to stop this.
- Inference from photos. The "Doxing via the Lens" study, accepted at ICLR 2026, built a benchmark of 500 real-world images, including private settings such as selfies. Across 11 multimodal models, AI consistently outperformed non-expert humans at geolocating images and could reveal sensitive locations such as home addresses or neighbourhoods. Models with web access and tool use were especially effective.
- Agent-based geolocation. Another research team built an AI agent that combines a vision model with search tools and showed it significantly increases location privacy risks, especially for difficult images that defeat simpler methods.
- Real-time identification of strangers. In October 2024, two Harvard students demonstrated a system that streamed video from commercially available smart glasses, sent detected faces to a public face search engine, and used a language model to combine the results with people-search data. Within seconds it could suggest a stranger's name, phone number and home address. The students did not release their code, and the system made errors, but the demonstration showed that every component was already off the shelf.
The common thread is that no single tool is the problem. The risk comes from chaining together capabilities that are each legal and easy to obtain.
Who Is Targeted
Anyone with an online presence can be doxed, but some groups face elevated risk:
- Journalists and researchers, whose reporting provokes retaliation from criminals, extremists or state-aligned groups.
- Activists and community organisers, particularly during protests or contested political moments.
- Executives and public figures, whose home addresses and family details are valuable for intimidation, extortion or physical attacks.
- Public servants, including police officers, judges, election workers and health officials.
- Women and minorities in public life, who are disproportionately targeted by coordinated harassment campaigns.
- Security researchers and moderators, who often operate under pseudonyms to protect themselves from the people they investigate.
- Ordinary users, who are increasingly doxed over online disputes, viral posts or gaming conflicts.
Why Organisations Should Care
AI doxing is not only a personal safety issue. It creates direct business risk:
- Executive protection. Profiles that combine an executive's home address, daily routine and family members enable physical threats, extortion and targeted kidnapping risk.
- Social engineering. Detailed personal profiles make phishing, voice-cloning fraud and business email compromise far more convincing. An attacker who knows an employee's manager, colleagues, hobbies and recent travel can craft very credible lures.
- Credential attacks. Personal details often answer security questions or appear in reused passwords.
- Staff safety and retention. Employees in visible roles, such as customer support, moderation, compliance or fraud investigation, may be targeted by angry customers or criminals.
- Reputational and legal exposure. Organisations may face scrutiny if their own systems, websites or data handling make employee or customer information easy to aggregate.
The Legal Landscape
Laws are catching up, but unevenly.
- Australia passed the Privacy and Other Legislation Amendment Act 2024, which created specific criminal offences for doxing. The aggravated offence, where people are targeted because of characteristics such as race, religion, sex, sexual orientation or disability, carries up to seven years' imprisonment.
- European Union data protection law gives individuals rights to access and erase personal data and restricts the processing of biometric data, which affects facial recognition services operating in Europe.
- Mauritius protects personal data under the Data Protection Act 2017, overseen by the Data Protection Office, and addresses online harassment and misuse of computer systems under the Cybersecurity and Cybercrime Act 2021. Victims can report incidents to the police and through CERT-MU's online cybercrime reporting system.
In many jurisdictions, however, collecting and combining publicly available information is not itself illegal. Legal protection usually applies only once information is used to harass, threaten or defraud, which means prevention remains essential.
How to Protect Yourself
No single step makes you invisible, but each one raises the cost of targeting you. The aim is to remove easy links between your online identities, your face and your physical location.
Reduce what you publish
- Review your public profiles on every platform and remove details you would not want a stranger to combine: employer, school, neighbourhood, daily routine, children's names and schools.
- Delete or archive old posts and forum accounts you no longer use. AI can read years of history in seconds.
- Avoid posting in real time from places you visit regularly. Share holiday or event photos after you have left.
Treat photos as location data
- Remove location metadata (EXIF) before sharing images. Most phones can disable location tagging in the camera settings.
- Check backgrounds before posting: house numbers, street signs, car number plates, distinctive views from windows, school uniforms and reflections can all reveal where you are.
- Be careful with photos taken at or around your home. Research shows AI can geolocate even indoor and private-setting images with worrying accuracy.
Break the links between identities
- Use different usernames on different platforms, especially for pseudonymous or sensitive accounts. Reused handles are one of the easiest ways to connect identities.
- Use separate email addresses and, where possible, separate phone numbers for public, personal and sensitive activities.
- Avoid using the same profile photo across professional and personal accounts. A single image can be matched across the web.
Remove data from people-search and face search services
- Search for your own name, phone number and email address to see what is publicly available.
- Use the opt-out or removal procedures offered by people-search sites and data brokers. Some services automate these requests for a fee.
- Some public face search engines offer opt-out requests that exclude your face from results. Use them, while recognising they do not remove the original images from the web.
- Where data protection law applies, use your legal right to request erasure of your personal data.
Secure your accounts
- Enable multi-factor authentication on email, social media and financial accounts, preferably with an authenticator app or security key rather than SMS.
- Use a password manager and unique passwords so that a breach on one site does not unlock others.
- Lock down account recovery options and do not use security questions whose answers can be found online.
Protect your household
- Talk to family members about what they post. A relative's public photo or tagged location can expose you just as easily as your own.
- Consider whether your home address appears in public registries or business filings, and use a business or postal address where that is legally permitted.
If You Have Been Doxed
- Prioritise safety. If you believe you are at risk of physical harm, contact the police immediately.
- Document everything. Take screenshots with dates, URLs and usernames before content is removed. This evidence matters for platform reports and legal action.
- Report the content. Most platforms prohibit publishing private information. Use their reporting tools and specify that personal data has been exposed.
- Secure your accounts. Change passwords, review recovery details and check for unfamiliar logins, as doxing often precedes account takeover attempts.
- Alert your bank and employer if financial details or workplace information were exposed, so they can watch for fraud and social engineering.
- Report it formally. In Mauritius, incidents can be reported to the police and to CERT-MU, and misuse of personal data can be raised with the Data Protection Office.
What Organisations Should Do
- Include doxing in threat models, especially for executives, high-profile staff and teams that deal with criminals or angry customers.
- Offer digital footprint reviews for at-risk employees, including data broker removal and social media guidance.
- Limit unnecessary exposure of staff details on corporate websites, press releases and public documents.
- Train staff on AI-enhanced social engineering, using realistic examples of how personal information is turned into convincing lures.
- Create a response procedure so that a doxed employee knows who to contact, how evidence will be preserved and what support is available.
Analyst Assessment
AI has turned doxing from a skilled, labour-intensive activity into a commodity capability. We assess with high confidence that:
- AI-assisted doxing will increase, because the tools are cheap, widely available and improving quickly, while the underlying data sources remain largely unchanged.
- Agentic AI will automate the full chain, from collecting posts to geolocating photos and enriching identities, making targeting at scale realistic for individuals as well as organised groups.
- Photos will become one of the most significant privacy risks, as geolocation and face matching models improve faster than most users' habits.
- Regulation will remain fragmented. A few countries are introducing specific doxing offences, but most rely on general harassment and data protection laws that apply only after harm occurs.
The most effective defence is to reduce the raw material. AI can only connect what it can find. People and organisations that limit what they publish, separate their identities, clean their photos and remove their data from aggregator sites make themselves dramatically harder to target, even as the tools improve.
Sources
- Beyond Memorization: Violating Privacy Via Inference with Large Language Models (ETH Zurich, ICLR 2024)
- Doxing via the Lens: Revealing Location-related Privacy Leakage on Multi-modal Large Reasoning Models (ICLR 2026)
- GEO-Detective: Unveiling Location Privacy Risks in Images with LLM Agents
- Someone Put Facial Recognition Tech onto Meta's Smart Glasses to Instantly Dox Strangers (404 Media)
- Privacy and Other Legislation Amendment Act 2024 (Federal Register of Legislation, Australia)
- Cybersecurity and Cybercrime Act 2021 (Mauritius, official text)
- Data Protection Office, Mauritius
- CERT-MU, Computer Emergency Response Team of Mauritius
- MAUCORS, the Mauritian Cybercrime Online Reporting System
