Definition
Plain language
A running map an AI agent keeps of where every piece of text it's handling originally came from.
As stated in the literature
In DASGuard, a per-step structure labeling content as trusted, clean, or untrusted and propagating suspicion forward, so that anything derived from untrusted text inherits its taint.
Why it matters: By remembering the origin of every piece of text, it lets an agent treat anything derived from untrusted sources with appropriate caution.
For example, if an agent copies text from an untrusted email into a new file, the graph marks that file as suspicious because of where its contents came from.
Heard on the show
“And it maintains, at every step, what the paper calls a content-source graph.”Episode 105 — The Trojan Is Your Agent's Memory: Why Single-Step Defenses Miss Persistent Attacks