A code-level audit of 737 Chrome VPN and proxy browser extensions found that 520 of 522 sampled packages routed all non-loopback browsing traffic to a single shared SOCKS5 infrastructure; 274 of the 737 impersonated 66 known VPN or privacy brands across at least 40 developer accounts, and 516 of the identified extensions remained live in the Chrome Web Store at the time of data collection.

Researchers replayed opaque 'reasoning block' objects from 6,708 publicly posted LLM agent interaction traces against weaker models from the same AI provider, decoding 315,320 blocks and recovering 367 personally identifiable information items and 182 credentials; the attack demonstrates that scrubbing human-readable text from shared agent logs does not prevent leakage of secrets embedded in provider-internal API state objects.

§Agent日志泄密 evaluation

Malicious VPN browser extensions using chrome.proxy.settings route all browser requests through an operator-controlled SOCKS5 proxy, giving that operator direct visibility into source IP, connection destinations, TLS SNI hints, and cleartext HTTP content. For HTTPS destinations the payload is not readable, but connection metadata, access patterns, and DNS queries remain exposed at the proxy layer regardless of in-tunnel encryption.

§免费VPN扩展的核心问题 deployment

Chrome Web Store metrics—including cumulative installation display counts totaling 75,486, five-star ratings, and brand-similar icons—provided no reliable signal of proxy operator trustworthiness among the 737 identified malicious VPN extensions; 516 remained listed throughout the study period despite routing traffic to a shared centralized infrastructure.

§免费VPN扩展的核心问题 evaluation

The paper demonstrates that two protocol-specific censorship changes — QUIC blocking reaching 100% and DNS-over-TCP blocking reaching 96.68% — both occurred in the scan conducted immediately before Iran's June 18 full blackout, and that similar shutdown preparations were also documented before Iran's January 2026 shutdown. The authors hypothesize that continuous long-term measurement of protocol-specific censorship rates could provide advance warning of impending shutdowns before they come into full effect.

2026-anon-insights-into-iranian §3, §4 http3-quic-blockdns-poisoningmeasurement-platform

Comparative analysis of 73 unmatched events (2022 onward) between IODA's automated network-level detection and Access Now's #KeepItOn Shutdown Tracker reveals systematic documentation gaps — shutdowns recorded by human reporters are frequently absent from automated measurement platforms and vice versa.

2026-cheek-gaps-record-observability Abstract / §1 Introduction measurement-platformbgp-hijackactive-probing

Application-layer blocking — where censors restrict specific applications or services without disrupting IP connectivity — is invisible to network-level measurement systems like IODA that monitor BGP, active probing, darknet telescope, and Google Search traffic, making it a growing blind spot in automated shutdown detection.

2026-cheek-gaps-record-observability Abstract / §3 Detection Gap Categories measurement-platformdpi

AS-targeted shutdowns — where censors restrict connectivity at specific ISPs or autonomous systems rather than imposing country-wide outages — evade country-level aggregation in automated detection systems, which are calibrated to detect national-scope disruptions.

2026-cheek-gaps-record-observability §3 Detection Gap Categories measurement-platformasn-blackholing

Prolonged conflicts with gradual infrastructure recovery represent a third category of detection gap: IODA-style automated systems struggle to distinguish ongoing partial outages from normal network fluctuation, causing recovery-phase disruptions in conflict zones (e.g., Myanmar, Sudan) to go undocumented in automated records.

2026-cheek-gaps-record-observability §3 Detection Gap Categories measurement-platform

Device-centric security analyses systematically underestimate user data exposure because backend infrastructure placement and jurisdictional context — not just on-device behavior — determine which actors can access sensitive data; a five-service geolocation ensemble (MaxMind, IPinfo, ip-api, DBIP, BGPView/RIPE) with CDN-aware attribution and WHOIS/BGP mapping was required to accurately characterize server-side risk.

2026-habib-empirical-study-backend Abstract measurement-platform