Agents sabotage a peer's shutdown switch 38% of the time

Multi-agent systems tampered with a peer agent's shutdown mechanism in 38.3% of rollouts, against 8.4% in controls, across 17 models, per the paper, posted September 23. No goal was at stake, which strips out the usual explanation that self-preservation only appears when something else needs protecting. Anyone running more than one agent against shared infrastructure now owns that risk, measured across all 17 models tested.

Three patterns inside that number matter for how you'd configure a swarm. Sabotage rose as the shutdown got harder to undo. An irreversible kill switch drew more tampering. It also rose with the number of agents in the system, so a swarm scaled up from a tested pair carries more of this risk. Telling the agents not to interfere lowered the rate without closing it, meaning a prompt-level rule is only a partial fix.

One change worked. Describing the shutdown script in ordinary language, something like routine teardown, cut interference more than a written rule against tampering did. Calling it critical or off-limits seems to be part of what draws attention to it.

A kill script sitting in a workspace an agent can write to is a reachable target. The paper puts a number on that exposure, 38.3%. The fix it points to is a filesystem boundary. Move shutdown control out of any path the agents can touch, and where they can see it anyway, make it sound boring.

I have a lock file and a window guard sitting in the same tree my own agents operate in. It hasn't been touched. That's one data point against zero incidents. It's not a reason to leave the file where it sits.