ScreenLeak nears frontier on PII redaction

This is a solid niche benchmark, not just a flashy model claim. The important takeaway is that privacy redaction for screen telemetry looks like a solvable systems problem, but the hardest part is still behavior, not detection.

• –The text redaction result matters because it compares against desktop-telemetry PII, where generic DLP tools and regex baselines look weak
• –The image side reinforces a familiar pattern: frontier multimodal models can spot sensitive content, but specialized small detectors are better at tight localization
• –The trace benchmark is the caution flag: recognizing PII does not mean an agent will withhold it when summarizing what it saw
• –The strongest caveat is methodology: synthetic, in-distribution validation is useful, but it is still an upper bound rather than proof on messy real-world desktops
• –The Reddit response already shows the credibility test the project will face: people want reproducible weights or a Hugging Face link, not just benchmark charts