2026-11-17 · America/Los_Angeles · 社区动态 · #19
I ran an automated attack-pattern miner on two agent frameworks. Three lessons on tool permissions and sandboxing
Disclosure: I’m the author of the write-up linked at the bottom, and this is my own research project. For the past two weeks I’ve been testing whether a static analysis pipeline can learn attack patterns by itself, without a human writing rules for each framework. The short version of how it works: Build a graph of the codebase: classes, method calls, data flow. Propose likely sources, sinks, guards and validators from that graph. Have a model filter out candidates that aren’t security-relevant, then check that real data flow exists. Turn the survivors into queries, run them against the…
热度 40.3 / 100;排名与评分保留该期记录。