Definition
Plain language
How many more steps an agent takes under attack compared to without the attack.
As stated in the literature
In LoopTrap, the ratio of attacked-trajectory step count to clean-trajectory step count for the same task, used to quantify slowdown caused by termination-poisoning attacks.
Also called: SAF
Why it matters: It quantifies how much an attacker can waste an agent's compute and time without ever causing a visible crash.
For example, if an agent normally finishes a task in 10 steps but takes 80 under attack, the step amplification factor is 8.
Heard on the show
“But it also means the SAF numbers are a clean-room upper bound.”Episode 030 — Why Your AI Agent Won't Stop Working — and Each Model Falls for a Different Trap