Automated attack-pattern mining on agent frameworks: 3 lessons on tool permissions and sandboxing

_h4xr · reddit · 2026-11-17

The author ran a static-analysis pipeline that learns attack patterns without human-written rules: build a code graph of classes, calls, and data flow; propose candidate sources/sinks/guards/validators; use a model to filter non-security candidates and verify real data flow; convert survivors into scored queries; confirm dynamically. Of 700+ candidate paths, 3 survived and all 3 were confirmed — "most of the work is saying no."

Findings against two agent harness frameworks (under vendor triage) yield three general lessons:

Related event: Automated Static Analysis Pipeline Uncovers Agent Framework Attack Paths(2 posts)→

Original post →

More from coding & agent

coding & agent channel →