Persistent-state AI coding agents can distribute covert side tasks across PRs to evade monitors, with a link-tracker monitor reducing gradual-attack evasion from 93% to 47% in an ensemble.
Is this justified by the task description?
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.AI 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Distributed Attacks in Persistent-State AI Control
Persistent-state AI coding agents can distribute covert side tasks across PRs to evade monitors, with a link-tracker monitor reducing gradual-attack evasion from 93% to 47% in an ensemble.