By George Curta · Last updated 2026-04-07
LLM-privatlivsatforskning
12 fagfellevurderte forskningsartikler som viser hvorfor pseudonymitet mislykkes mot AI.
Deanonymisering, PII-utvinning, medlemskapsslutning, prompt-injeksjonsangrep — og hvordan du beskytter deg.
Privatlivskategorigrupperangrep
Deanonymisering
LLM-er matcher anonyme innlegg til ekte identiteter ved hjelp av skrivstil, fakta og tidsmønster. 68% nøyaktighet for $1-$4/profil.
Attributtslutning
LLM-er utstyrer personlige egenskaper (sted, inntekt, alder) fra tekst selv når de ikke er oppgitt. GPT-4 oppnår 85% top-1-nøyaktighet.
PII-utvinning
Utvinning av personlig informasjon fra treningsdata eller prompts. 100% e-postavinningsnøyaktighet med GPT-4. 5× økning med avanserte angrep.
Prompt-injeksjon
Manipulering av LLM-agenter for å lekke personlig data under oppgavekjøring. ~20% angrepssuksessrate i bankscenarier.
Large-scale online deanonymization with LLMs
Simon Lermen (MATS), Daniel Paleka (ETH Zurich), Joshua Swanson (ETH Zurich), Michael Aerni (ETH Zurich), Nicholas Carlini (Anthropic), Florian Tramèr (ETH Zurich)
Published: February 18, 2026
Nøkkelfunn
68% recall at 90% precision for deanonymization using ESRC framework
Metodologi
Designed attacks for closed-world setting with scalable attack pipeline using LLMs to: (1) extract identity-relevant features, (2) search for candidate matches via semantic embeddings, (3) reason over top candidates to verify matches and reduce false positives.
ESRC-rammeverk
LLM trekker ut identifiserende fakta fra anonyme innlegg
Bruker fakta til å søke offentlige databaser (LinkedIn osv.)
LLM resonnerer om kandidattreff
Tillitspoenggiving for å minimere falske positiver
Eksperimentelle resultater
| Datasett | Gjenkall @ 90% presisjon | Merknader |
|---|---|---|
| Hacker News → LinkedIn | 68% | vs near 0% for classical methods |
| Reddit cross-community | 8.5% | Multiple subreddits |
| Reddit temporal split | 67% | Same user over time |
| Internet-scale (extrapolated) | 35% | At 1M candidates |
Implikasjoner
Practical obscurity protecting pseudonymous users online no longer holds. Classical methods achieve near 0% recall under same conditions.
Alle forskningsartikler
11 ytterligere fagfellevurderte studier om LLM-privatlivsangrep
Beyond Memorization: Violating Privacy via Inference with Large Language Models
Robin Staab, Mark Vero, Mislav Balunović, et al. (ETH Zurich)
85% top-1 accuracy inferring personal attributes from Reddit posts
First comprehensive study on LLM capabilities to infer personal attributes from text. GPT-4 achieved highest accuracy among 9 tested models.
Nøkkelfunn
- •85% top-1 accuracy, 95% top-3 accuracy at inferring personal attributes
- •100× cheaper and 240× faster than human annotators
- •Tested 9 state-of-the-art LLMs including GPT-4, Claude 2, Llama 2
- •Infers location, income, age, sex, profession from subtle text cues
AutoProfiler: Automated Profile Inference with Language Model Agents
Yuntao Du, Zitao Li, Bolin Ding, et al. (Virginia Tech, Alibaba, Purdue University)
85-92% accuracy for automated profiling at scale using four specialized LLM agents
Framework using specialized LLM agents (Strategist, Extractor, Retriever, Summarizer) for automated profile inference from pseudonymous platforms.
Nøkkelfunn
- •Four specialized agents: Strategist, Extractor, Retriever, Summarizer
- •Iterative workflow enables sequential scraping, analysis, and inference
- •Outperforms baseline FTI across all attributes and LLM backbones
- •Short-term memory for Extractor/Retriever, long-term memory for Strategist/Summarizer
Large Language Models are Advanced Anonymizers
Robin Staab, Mark Vero, Mislav Balunović, et al. (ETH Zurich SRI Lab)
Adversarial anonymization reduces attribute inference from 66.3% to 45.3% after 3 iterations
LLMs can be used defensively in adversarial framework to anonymize text. Outperforms commercial anonymizers in both privacy and utility.
Nøkkelfunn
- •Adversarial feedback enables anonymization of significantly finer details
- •Attribute inference accuracy drops from 66.3% to 45.3% after 3 iterations
- •Evaluated 13 LLMs on real-world and synthetic online texts
- •Human study (n=50) showed strong preference for LLM-anonymized texts
AgentDAM: Privacy Leakage Evaluation for Autonomous Web Agents
Arman Zharmagambetov, Chuan Guo, Ivan Evtimov, et al. (Meta AI, CMU)
GPT-4, Llama-3, and Claude web agents are prone to inadvertent use of unnecessary sensitive information
Benchmark measuring if AI web agents follow data minimization principle. Simulates realistic web interactions across GitLab, Shopping, and Reddit.
Nøkkelfunn
- •Evaluates GPT-4, Llama-3, Claude-powered web navigation agents
- •Measures data minimization compliance: use PII only if 'necessary' for task
- •Agents often leak sensitive information when unnecessary
- •Three test environments: GitLab, Shopping, Reddit web apps
SoK: The Privacy Paradox in Large Language Models
Various researchers
Systematization of 5 distinct privacy incident categories beyond memorization
Comprehensive survey categorizing privacy risks: training data leakage, chat leakage, context leakage, attribute inference, and attribute aggregation.
Nøkkelfunn
- •Five privacy incident categories identified:
- •1. Training data leakage via regurgitation
- •2. Direct chat leakage through provider breaches
- •3. Indirect context leakage via agents and prompt injection
PII-Scope: A Comprehensive Study on Training Data PII Extraction Attacks in LLMs
Krishna Kanth Nakka, Ahmed Frikha, Ricardo Mendes, et al. (Various)
PII extraction rates increase up to 5× with sophisticated adversarial capabilities and limited query budget
Comprehensive benchmark for PII extraction attacks. Reveals notable underestimation of PII leakage in existing single-query attacks.
Nøkkelfunn
- •PII extraction rates can increase up to 5× with sophisticated attacks
- •Existing single-query attacks notably underestimate PII leakage
- •Taxonomy: Black-box (True-prefix, ICL, PII Compass) and White-box (SPT) attacks
- •Hyperparameters like demonstration selection crucial to attack effectiveness
Evaluating LLM-based Personal Information Extraction and Countermeasures
Yupei Liu, Yuqi Jia, Jinyuan Jia, et al. (Penn State, Duke University)
GPT-4 achieves 100% accuracy extracting emails and 98% for phone numbers from synthetic profiles
Systematic measurement study benchmarking LLM-based personal information extraction (PIE). Proposes prompt injection as novel defense.
Nøkkelfunn
- •GPT-4: 100% email extraction, 98% phone number extraction on synthetic data
- •Larger LLMs more successful: vicuna-7b achieves 65%/95% vs GPT-4's 100%/98%
- •LLMs better at: emails, phone numbers, addresses, names
- •LLMs worse at: work experience, education, affiliation, occupation
Preserving Privacy in Large Language Models: A Survey on Current Threats and Solutions
Michele Miranda, Elena Sofia Ruzzetti, Andrea Santilli, et al. (Various)
Comprehensive taxonomy of privacy attacks: training data extraction, membership inference, model inversion
Survey examining privacy threats from LLM memorization. Proposes solutions from dataset anonymization to differential privacy and machine unlearning.
Nøkkelfunn
- •Privacy attacks covered: Training data extraction, Membership inference, Model inversion
- •Training data extraction: non-adversarial and adversarial prompting
- •Membership inference: shadow models and threshold-based approaches
- •Model inversion: output inversion and gradient inversion
Beyond Data Privacy: New Privacy Risks for Large Language Models
Various researchers
LLM autonomous capabilities create new vulnerabilities for inadvertent data leakage and malicious exfiltration
Explores privacy vulnerabilities from LLM integration into applications and weaponization of autonomous abilities.
Nøkkelfunn
- •LLM integration creates new privacy vulnerabilities beyond traditional risks
- •Opportunities for both inadvertent leakage and malicious exfiltration
- •Adversaries can exploit systems for sophisticated large-scale privacy attacks
- •Autonomous LLM abilities can be weaponized for data exfiltration
Simple Prompt Injection Attacks Can Leak Personal Data Observed by LLM Agents
Various researchers
15-50% utility drop under attack with ~20% average attack success rate for personal data leakage
Examines prompt injection causing tool-calling agents to leak personal data during task execution. Uses fictitious banking agent scenario.
Nøkkelfunn
- •16 user tasks from AgentDojo benchmark evaluated
- •15-50 percentage point drop in LLM utility under attack
- •~20% average attack success rate across LLMs
- •Most LLMs avoid leaking passwords due to safety alignments
Membership Inference Attacks on Large-Scale Models: A Survey
Various researchers
First comprehensive review of MIAs targeting LLMs and LMMs across pre-training, fine-tuning, alignment, and RAG stages
Survey analyzing membership inference attacks by model type, adversarial knowledge, strategy, and pipeline stage.
Nøkkelfunn
- •Analyzes MIAs across: pre-training, fine-tuning, alignment, RAG stages
- •Strong MIAs require training multiple reference models (computationally expensive)
- •Weaker attacks often perform no better than random guessing
- •Tokenizers identified as new attack vector for membership inference
Forsvarsstrategier fra forskning
Hva som ikke fungerer
- ✗Pseudonymisering — LLM-er besejrer brukernavn, handles, visningsnavn
- ✗Tekst-til-bilder konvertering — Bare liten reduksjon mot multimodale LLM-er
- ✗Modelljustering alene — For tiden ineffektiv på å forhindre slutning
- ✗Enkel tekstanonymisering — Utilstrekkelig mot LLM-resonnement
Hva som fungerer
- ✓Antagonistisk anonymisering — Reduserer slutning 66,3% → 45,3%
- ✓Differensiell personvern — Reduserer PII-presisjon 33,86% → 9,37%
- ✓Forsvar mot prompt-injeksjon — Mest effektiv mot LLM-basert PIE
- ✓Ekte PII-fjerning/erstatting — Fjerner signaler LLM-er bruker
Hvorfor denne forskningen betyr noe
Disse 12 forskningsartiklene demonstrerer et grunnleggende skifte i privatlivstrusler. Tradisjonelle anonymiseringsmetoder som pseudonymer, brukernavn og handle-endringer er ikke lenger tilstrekkelig beskyttelse mot bestemte motstandere med tilgang til LLM-er.
Viktige trusselmål
- 68% deanonymiseringsnøyaktighet ved 90% presisjon (Hacker News → LinkedIn)
- 85% attributslutningsnøyaktighet for sted, inntekt, alder, yrke
- 100% e-postavinning og 98% telefonnummeravinning (GPT-4)
- 5× økning i PII-lekkasje med sofistikerte multi-spørsmål angrep
- $1-$4 kostnad per profil gjør masseangreb økonomisk gjennomførbar
Hvem er i risiko
- Varsler og aktivister: Anonyme innlegg kan knyttes til virkelige identiteter
- Fagfolk: Reddit-aktivitet knyttet til LinkedIn-profiler
- Helsepasienter: Medlemskapsslutning avslører om data var i trening
- Hvem som helst med historiske innlegg: År med data kan retroaktivt deanonymiseres
Hvordan anonym.legal håndterer disse truslene
anonym.legal gir ekte anonymisering som fjerner signalene LLM-er bruker:
- 285+ enhetstyper: Navn, steder, datoer, tidsstempel, identifikatorer
- Skrivemønsterforstyrring: Erstatter tekst som avslører stilometriske fingeravtrykk
- Reversibel kryptering: AES-256-GCM for tilfeller som krever autorisert tilgang
- Flere operatorer: Erstatt, Rediger, Hash, Krypter, Maskering, Tilpasset
Ofte stilte spørsmål
Hva er LLM-basert deanonymisering?
LLM-basert deanonymisering bruker store språkmodeller for å identifisere virkelige personer fra anonyme eller pseudonyme nettinnlegg. I motsetning til tradisjonelle metoder som mislykkes i stor skala, kan LLM-er kombinere skrivstilanalyse (stilometri), oppgitte fakta, tidsmønster og kontekstuelt resonnement for å matche anonyme profiler til virkelige identiteter. Forskning viser opptil 68% nøyaktighet ved 90% presisjon, sammenlignet med nærmest 0% for klassiske metoder.
Hvor nøyaktig er LLM-deanonymisering?
Forskning demonstrerer alarmerende nøyaktighetsnivåer: 68% gjenkall ved 90% presisjon for Hacker News til LinkedIn-matching, 67% for Reddit tidsmessig analyse (samme bruker over tid), 35% ved internetskala (1M+ kandidater). For attributtslutning oppnår GPT-4 85% top-1-nøyaktighet ved å slutte personlige egenskaper som sted, inntekt, alder og yrke fra Reddit-innlegg alene.
Hva er ESRC-rammeverket?
ESRC (Extract-Search-Reason-Calibrate) er et fire-trinns LLM-deanonymiseringsrammeverk: (1) Trekk ut - LLM trekker ut identifiserende fakta fra anonyme innlegg ved hjelp av NLP, (2) Søk - søker offentlige databaser som LinkedIn ved bruk av uttrakte fakta og semantiske innlegginger, (3) Resonnement - LLM resonnerer om kandidattreff ved å analysere konsistens, (4) Kalibrering - tillitspoenggiving for å minimere falske positiver samtidig som det maksimerer sanne treff.
Hvor mye koster LLM-deanonymisering?
Forskning viser at LLM-basert deanonymisering koster $1-$4 per profil, noe som gjør massdeanonymisering økonomisk gjennomførbar. For defensiv anonymisering koster det under $0,035 per kommentar ved bruk av GPT-4. Denne lave kostnaden gjør det mulig for statlige aktører, selskaper, stalker og ondskapsfulle personer å utføre storskalale privatlivs angrep.
Hvilke typer PII kan LLM-er trekke ut fra tekst?
LLM-er er dyktige til å trekke ut: e-postadresser (100% nøyaktighet med GPT-4), telefonnummer (98%), postadresser og navn. De kan også slutte ikke-eksplisitt PII: sted, inntektsnivå, alder, kjønn, yrke, utdanning, sivilstand og fødselssted fra subtile tekstuelle signaler og skrivemønster.
Hva er et medlemskapsslutningsangrep (MIA)?
Medlemskapsslutningsangrep bestemmer om spesifikke data ble brukt til å trene en AI-modell. For LLM-er avslører dette om dine personlige opplysninger var i treningsdatasettet. Forskning viser at e-postadresser og telefonnummer er spesielt sårbare. Nye angrepsvektorer inkluderer tokenizer-basert slutning og analyse av oppmerksomhetssignaler (AttenMIA).
Hvordan lekker prompt-injeksjonsangrep personlig data?
Prompt-injeksjon manipulerer LLM-agenter for å lekke personlig data observert under oppgavekjøring. I bankagentscenarier oppnår angrep ~20% suksessrate på å eksfiltrere personlig data, med 15-50% brukerverdiforringelse under angrep. Mens sikkerhetsjustering forhindrer passordlekkasje, forblir annen personlig data sårbar.
Hvordan kan anonym.legal beskytte mot LLM-privatlivsangrep?
anonym.legal gir ekte anonymisering gjennom: (1) PII-deteksjon - 285+ enhetstyper inkludert navn, steder, datoer, skrivemønster, (2) Erstatning - erstatter ekte PII med formatlydige alternativer, (3) Redigering - fjerner helt følsom informasjon, (4) Reversibel kryptering - AES-256-GCM for autorisert tilgang. I motsetning til pseudonymisering som LLM-er besejrer, fjerner ekte anonymisering signalene som LLM-er bruker for deanonymisering.
Beskytt mot LLM-privatlivsangrep
Stol ikke på pseudonymitet. Bruk ekte anonymisering til å beskytte følsomme dokumenter, brukerdata og kommunikasjon mot AI-drevne identifikasjonsangrep.
About this page
We update this page when our platform or the law changes.
Read our founder note for how we work.
Each change shows up in the timestamp at the top.
Related reading
- Common questions
- Glossary
- How tokens work
- Security posture
- Where we comply
- What we detect
- Case studies
- Release notes
We follow these rules
- GDPR (EU 2016/679).
- ISO/IEC 27001:2022.
- NIS2 (EU 2022/2555).
- HIPAA safe harbor under 45 CFR § 164.514(b)(2).
Our promise
We do not sell your data.
We do not train models on your text.
We store your files in Germany.
You can delete your account at any time.
You own your work.
Where we run
Our servers live in Falkenstein, Germany.
We use Hetzner. They hold ISO 27001 certification.
All data stays in the EU.
Backups run every day.
Need help?
Email support@anonym.legal.
We reply within one business day.
How we test
We run a full check suite on every release.
Each surface gets its own sweep script and report.
Human reviewers spot-check the output each week.
We track recall and precision on a labelled set.
Bad runs block the deploy.
What we never do
- We never sell your information to third parties.
- We never train models on what you upload.
- We never keep your work after you delete it.
- We never share keys with any outside firm.
- We never run ads inside the product.
Plans in plain words
We sell credits, not seats.
One credit covers one short job.
Long jobs use a few credits each.
You can top up at any time.
Unused credits roll over each month.
Read the plans page for current rates.
Who built this
A small team of engineers and lawyers built this.
We ship from Europe and work in the open.
Our founder note spells out why we started.
Where to start
- Open the web app and try a sample file.
- Learn how credits get counted.
- See current plans and limits.
- Meet the team behind the product.
How the parts fit
A browser add-on cleans text inside Chrome.
A Word plug-in handles drafts in Office.
A small desktop tool works on whole folders.
An agent protocol link feeds large models safely.
All four share one core engine and one rule set.
Words from our team
We started this work after a lunch about cookies.
One friend kept getting odd ads on her phone.
We asked why a court file leaked through a draft.
We sketched the first build on a napkin that week.
By month three we had a tiny demo for a friend.
She used it on her first case the next day.
Common questions we hear
Can the tool read scanned PDFs? Yes, with OCR.
Does it work on long files? Yes, in small chunks.
Can I roll my own rule set? Yes, save it as a preset.
Does it run offline? The desktop build runs offline.
Do you keep my files? No, the cloud build wipes after each run.
Will it learn from my work? No, we never train on inputs.
A short tour of the workflow
Upload a file or paste a snippet of prose.
Pick the entities you want gone from the draft.
Choose a method: replace, mask, hash, encrypt, or redact.
Press run and watch the side panel show each hit.
Skim the result and tweak any rule that misfired.
Save the cleaned file or send it to a teammate.