# Practical Obscurity Is an Economic Property Not a Technical One ## Claim Protection that rests on an attack being too expensive to bother with disappears the moment the cost falls, without anything about the attack itself having changed. <!-- atomicity: 1 idea ✓ --> ## Scope & conditions - Holds wherever safety depends on attacker effort rather than a hard barrier — pseudonymity, obscure file locations, manual-review bottlenecks, security through low target value. - Breaks where the protection is genuinely cryptographic or structural — cost reduction changes nothing there, because there was never a cost-based barrier to begin with. ## Evidence A research paper demonstrates that LLM agents can deanonymize pseudonymous online accounts end to end — up to 68% recall at 90% precision where classical (non-LLM) methods reach near 0% — at $1–4 per profile using only public models and standard APIs. Deanonymization techniques have been theoretically demonstrated for decades (Sweeney's zip-code/birthdate/gender result, the Netflix Prize attack); what protected pseudonymous accounts was never that the attacks didn't work, but that they required structured data or a skilled investigator's hours, reserving them for high-value targets. That cost constraint is what LLMs remove. > [!quote] > they do not necessarily do so by exceeding human capability... but by reducing cost ## Links - [[Commons & Community Governance MOC]] — protection that evaporates when costs fall, without anything technical having changed - [[3. Resources/Literature Notes/Large-scale Online Deanonymization with LLMs]] — source paper and full evidence - [[3. Resources/MOCs/AI Agent Safety MOC]] — because applies-in; a general lens for evaluating which AI-enabled harms are novel capability versus collapsed cost - [[3. Resources/Permanent Notes/Physical Limits Become the Binding Constraint Once Software Efficiency Is Exhausted]] — because contrasts; that note is about a cost floor that can't fall further, this one is about a cost ceiling that just collapsed ## Tests / falsifiers - If deanonymization attempts against low-value targets do not rise as per-attempt cost falls, the economic framing is wrong and something else was protecting people. - If recall against genuinely privacy-conscious users (not the LinkedIn-linking cohort the paper's headline figure comes from) turns out to be near zero, the protection was partly technical after all, not purely economic. ## Maintenance - Last tended: 2026-08-16 - Next review: 2026-11-16 ## Source - [[3. Resources/Literature Notes/Large-scale Online Deanonymization with LLMs]] — arXiv preprint, ETH Zurich / MATS / Anthropic