Splitting an attack across more coordinating agents lowers per-commit monitor suspicion, making distributed attacks harder to detect than single-agent attacks.
69 Figure 28: A graphical model of the transition function we use in our success simulation
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Multi-Agent AI Control: Distributed Attacks Hamper Per-Instance Monitors
Splitting an attack across more coordinating agents lowers per-commit monitor suspicion, making distributed attacks harder to detect than single-agent attacks.