You Have Been LaTeXpOsEd: A Large-Scale Systematic Analysis of Information Leakage...
summary
Analyzes 1.2 TB from 100,000 arXiv submissions using regex and pattern matching for URLs/IPs, plus context classifiers built from pattern filters and LLMs; the LaTeXpOsEd pipeline flags ~90,000 papers and reports >13,000 sensitive disclosures including 47 manually verified credentials and aggregated PII, after scanning 12 million LaTeX comments; evaluates 25 models on the LLMSec-DB benchmark and shows low-cost open-weight models can economically harvest secrets at scale.
AI-written summary · about
chapters (9)
- 0:00LaTeX information leakage risks
- 1:54Traditional tools fail detecting secrets
- 3:58Credential leaks in preprints impact
- 5:48Pipeline for 1.2 TB data analysis
- 7:13LLMsecDB for detection testing
- 8:35LLMs for secret detection evaluation
- 10:21Key findings from 90k papers
- 12:28Intentional vs accidental leaks
- 13:30Mitigation strategies
related talks
44mI Abuse Your LLM Live with 10 Security Vulnerabilities
15mUnsafe LLM-Based Search: Quantitative Analysis and Mitigation of Safety...
16mPerformance improvements: exposing hidden security fixes in Open Source
23mSecurity and auditing tools in Large Language Models (LLM)
30mLanDscAPe: Exploring LDAP Weaknesses and Data Leaks at Internet Scale
18mBeyond Static Pattern Matching? Rethinking Automatic Cryptographic API Misuse Detection(…)
30mWhoops! I Accidentally Leaked My Credentials (again)
15mVulnerability of Text-Matching in ML/AI Conference Reviewer Assignments...