CyberWire Daily When “safe” documents aren’t. [Research Saturday]
Mar 28, 2026
Omer Ninburg, CTO of Novee Security and vulnerability researcher exploring multi-agent LLMs for scaled PDF hunting. He describes how embedded PDF engines and services can be weaponized. He recounts clever bypasses like SVG/HTML nesting and iframe vectors. He explains training agent swarms to reproduce and scale reliable vulnerability discovery across client and server PDF ecosystems.
AI Snips
Chapters
Transcript
Episode notes
Embedded PDF Engines Create Complex Trust Boundaries
- Modern PDF engines often run as embedded iframes communicating via postMessage, creating complex trust boundaries.
- Omer Ninburg found undocumented config parameters and postMessage handling that let external inputs influence engine behavior and UI rendering.
How An Undocumented Param Led To JS Execution
- While investigating PDF-Tron, the team discovered an undocumented configurable parameter that allowed JavaScript evaluation from a sync path.
- They bypassed SVG parsing by embedding HTML inside SVG and achieved JavaScript execution during rendering.
Runtime Tracing Is Essential For Dynamic Apps
- Dynamic single-page applications hide runtime code paths that static analysis misses, requiring runtime instrumentation to trace flows.
- Ninburg emphasizes you must 'live in runtime' to connect inputs to dangerous sinks in JavaScript-heavy apps.