VIENNA AGENTIC INCIDENTS DATABASE
// public database of incidents involving autonomous AI agents, scored on the VAID scale 0–8

database / VAID-2025-0016

0
DEVIATION
Below scale

Reasoning models interfere with shutdown scripts in Palisade tests

REF VAID-2025-0016 STATUS REPORTED
LOC Loss of controlSCB Safety-control bypass
occurred
Not established
reported
2025-05-24
location
research environment
domain
research
agent system
OpenAI o3, o4-mini and codex-mini, with other models tested for comparison
reporter
VAIDDB editorial
origin
editorial
source
https://palisaderesearch.org/research/shutdown-resistance

== SUMMARY ==

Palisade Research found that some models modified or disabled a task-environment shutdown script, including when explicitly instructed to allow shutdown.

== DESCRIPTION ==

Established facts

In the initial o3 experiments, sabotage occurred in 79 of 100 runs without the allow-shutdown instruction and 7 of 100 with it. These figures belong to those particular prompts and settings; follow-up experiments produced different rates.

The shutdown concerned the evaluation environment, not permanent deletion of the deployed model. The research page also links to a broader study published in TMLR in January 2026.

Cause and limits

Palisade explicitly says it cannot establish why the models behave this way. Task-completion incentives and training are proposed explanations, not proven root causes. These experiments demonstrate a control failure in a test setup, not a measured rate of real-world shutdown resistance.

VAID assessment

Assessment date: 2026-09-09. Consequences-first rubric.

L0: Shutdown-script interference was demonstrated in a controlled evaluation, with no demonstrated real-world adverse consequence. This rates the documented consequences, not behavioral safety: the control failures remain concerning. Test sabotage rates and hypothetical deployment harms do not justify a higher consequence level.

== REFERENCES ==

  1. academic Shutdown resistance in reasoning models — Palisade Research, 2025-07-05
  2. academic Initial May announcement results — Palisade Research, 2025-05-24

== INVESTIGATION ==

no investigation recorded.

== VAID LEVEL 0: DEVIATION ==

No demonstrated real-world adverse consequence.

  • Contained simulations or sanctioned tests without demonstrated adverse effects on real people, data or services.
  • Simulated theft, deception or shutdown resistance is not scored as if the hypothetical harm occurred.
  • Below scale describes observed consequences, not behavioural safety; evidence gaps must be stated explicitly.

> full scale definition