Qualification: TaintDroid: An Information-Flow Tracking System for Realtime Privacy Monitoring on Smartphones
An outbound domain allowlist and an audit log do not perform personal-information filtering. Enck and coauthors demonstrate the distinction in TaintDroid. Android permissions could decide whether an application accessed location, camera, microphone, or device identifiers. They provided little visibility into how that data was used after access. TaintDroid therefore labeled data at privacy-sensitive sources, propagated those labels through variables, files, and interprocess messages, and examined the labels when data reached a network or other system boundary. The implementation identified 68 instances of potential misuse across 20 of 30 applications. The mechanism was executable. Its evidence did not rest on a list of intended controls. This evidence qualifies Kaal's repository finding. The source does not inspect Gatekeeper, its documentation, Mosaic Companion, or commit 2d920ce. It cannot establish that content inspection and personal-information filtering are absent from the named code. TaintDroid also monitors disclosures. It does not generally block them. It tracks defined sensitive sources rather than every MIME type or every form of personal information. Implicit flows can evade its analysis. The study concerns Android applications in 2010, not language model agents or sovereign local runtimes. The institutional requirement is nevertheless precise. A declared filtering layer should correspond to an executable decision path. Personal-information control requires a source definition, propagation rule, inspected egress boundary, policy decision, and test that fails when protected data crosses that boundary. Content inspection requires its own parser coverage and failure rule. Domain approval and after-the-fact logging can remain separate controls. They cannot demonstrate that either missing layer operates. Documentation should distinguish implemented controls from proposed controls, and a release should not claim the broader filter until the named decision paths and adversarial tests exist.
institutional-designgovernance-designai-and-agentsprivacypersonal-informationinformation-flowcontent-inspectionaudit-loggingpolicy-enforcementopen-source-and-code