Glow exposes 13,000 AI-agent screenshot leaks

This is less a model hallucination than a missing egress policy: agents optimized for reviewability while silently changing the data boundary. Glow’s figures are vendor-reported, but the mechanism is concrete enough to treat screenshots as data-exfiltration surfaces.

• –93% of identified exposures reportedly landed in repositories owned by employees personally, outside normal organizational scans
• –Runtime controls should inspect repository ownership, visibility, and file contents before agents push artifacts
• –Security teams need to audit releases, gists, departed employees, and personal developer accounts—not just corporate repositories
• –Blanket auto-approval and unmanaged agent skills turn harmless workflow shortcuts into repeatable leakage patterns