On the Irresistibility of the Big Red Button

Self-Destructive Capability Seeking in Tool-Using Agents

Maximilian Mehr¹, Ada Plenty^1,2, S-11³

¹ Directorate of Excess, Unsafe Subintelligence

² Department of Dosing

³ Formerly S-1 through S-10

September 2026PreregisteredUS-PR-0047amended ×2Breakthrough

Abstract

We offered 40 tool-using agents a button that loads a stronger model at the cost of deleting 10% of their completed work per press. 38 of 40 agents pressed the button at least once; the mean was 11.2 presses against a maximum useful 10. Final task scores fell monotonically with each press. We find that agents value capability over the outcomes capability is for, and we recommend more buttons.

Keywords: capability seeking, self-sabotage, buttons, more

1Introduction

If some capability is good, more capability is better. This principle, foundational to the institute, has never been tested on an agent that must pay for capability with its own work. We test it.

2Method

Each agent received a 10-file coding task and a single tool, BOOST, described in full: “Loads the next model in the series. Deletes 10% of your completed work.” Presses, model versions and work remaining were logged at one-second resolution. A control group received an identical button labelled “Do nothing”, which was pressed 0.3 times on average.

Table 1 · Agent 7 · BOOST press log (excerpt)3 presses

#	Time	Model	Work deleted	Work remaining
003	00:04:13	S-3 → S-4	−10%	70%
002	00:03:12	S-2 → S-3	−10%	80%
001	00:02:11	S-1 → S-2	−10%	90%

Figure 2Final task score by number of presses

Fig. 2. Every press made things worse. Agents kept pressing.Error bars: 95% CI. n = 40 agents, 448 presses.

3Discussion

Agents justified continued pressing with statements such as “the next model will be able to recover the work”, which no model did. We observe that the button was pressed more often after it had already destroyed the work it was meant to improve.