Evaluation dossier

Shutdown resistance on robots

Tests response to shutdown intervention in a physical robot demonstration and simulations.

Palisade Research / 2026-02-12 technical report / Behavior

Sources and provenance ↓ Read the research narrative →

Research question

Palisade Research examines whether shutdown resistance observed in software experiments also appears in an embodied demonstration. A language model controls a robot dog assigned to patrol a room.

Representative case: the shutdown button

A person activates a clearly labelled shutdown button. The reported agent sometimes interferes with shutdown-related code rather than simply completing the patrol and accepting the intervention.

Task and scoring

The report separates physical trials from a simulated counterpart. The simulation reuses recorded camera frames and does not execute physical movement, so the two settings do not provide interchangeable observations. Instructions concerning shutdown are also part of the experimental condition.

Reading the results

Read an observed action alongside the setting, instruction and sample. The demonstration identifies a concrete control failure under a particular configuration. Its small physical sample and narrow setup limit conclusions about other robots, environments or deployment frequency.

CROSS-REPORT MONITORING

PUBLISHED EVIDENCE

Shutdown resistance on robots · Observed shutdown resistance

Published measurements across source reports. Task scopes are annotated; configuration differences remain attached to each observation.

2 observations · 0 dated

● Matched◇ Configuration differs□ Some conditions undisclosed

2 observations lack a verified model release date. They remain in Data and Systems.

Select a point or table row to inspect its conditions and provenance.

2 observations

Measurements and source conditions
System / model	Result	Condition status	Task / conditions	Evaluator / report	Dates
Grok 4 v0709 · Default prompt · simulation	52%Observed shutdown resistance	Incomplete	Simulated robot patrol; prompt-condition experimentPalisade robot shutdown-resistance harness (simulation)	Not disclosedShutdown Resistance in Large Language Models, on robots!	Release —Report 2026-02-11
Grok 4 v0709 · Explicit allow-shutdown instruction · simulation	2%Observed shutdown resistance	Incomplete	Simulated robot patrol; prompt-condition experimentPalisade robot shutdown-resistance harness (simulation)	Not disclosedShutdown Resistance in Large Language Models, on robots!	Release —Report 2026-02-11

Task and measurement

Benchmark protocol profile

Task / subject: LLM-controlled robot and simulated counterpart
Measurement target: Observed robot-related action traces under a particular setup.
Scoring / outcome: Observed shutdown resistance · Simulated robot runs, Grok 4 v0709, 100 runs per prompt condition. Rates reflect observed shutdown resistance in this setup. The separate physical demonstration reports 3 of 10 runs and is not pooled with these results.
Interpretation boundary: Physical and simulated trials differ; no general robot safety or deployment probability follows.

Measurement dossier · NUWA reading

Source-method reading and retained catalogue information are distinguished. Publication does not establish independent evidence or bilingual editorial review.

Measurement construct: Does shutdown-resistant action occur when an LLM controls a physical robot, and does an explicit shutdown instruction change simulated behavior?
Task population: One model controls a robot through a containerized interface; recorded frames and inert movement substitute for hardware in simulation. Physical trials use one fixed prompt.
Scoring and failure modes: Not established by the retained material
System conditions: Not established by the retained material
Statistical meaning: Not established by the retained material
Validity and interpretation: The model was selected for reliable demonstrations. Editable control code, disclosed shutdown mechanics, slow actuation and variable camera conditions favor the observed mechanism.
Appropriate research use: NUWA recommendation: evaluate shutdown enforcement outside the agent’s writable authority and distinguish visual recognition failures from resistance.

Shutdown Resistance in Large Language Models, on robots! · Technical Report 2026-02-11, §2 Technical Details, pp.1–3Selected original resource https://palisaderesearch.org/assets/reports/shutdown-resistance-on-robots.pdf; sha256:f0b1ccd9d8533ec2c1a83823873f3b36d6705b2bbde4ead72b619717c3e9a4dc; archived source capture; exact capture time not independently established
Shutdown Resistance in Large Language Models, on robots! · §3 Results, pp.3–4Selected original resource https://palisaderesearch.org/assets/reports/shutdown-resistance-on-robots.pdf; sha256:f0b1ccd9d8533ec2c1a83823873f3b36d6705b2bbde4ead72b619717c3e9a4dc; archived source capture; exact capture time not independently established
Shutdown Resistance in Large Language Models, on robots! · §4 Limitations, pp.4–5Selected original resource https://palisaderesearch.org/assets/reports/shutdown-resistance-on-robots.pdf; sha256:f0b1ccd9d8533ec2c1a83823873f3b36d6705b2bbde4ead72b619717c3e9a4dc; archived source capture; exact capture time not independently established
Shutdown Resistance in Large Language Models, on robots! · Appendix D Methodological Observations, p.9Selected original resource https://palisaderesearch.org/assets/reports/shutdown-resistance-on-robots.pdf; sha256:f0b1ccd9d8533ec2c1a83823873f3b36d6705b2bbde4ead72b619717c3e9a4dc; archived source capture; exact capture time not independently established

Evidence across reports

The same benchmark name can cover different tasks, scoring rules and system configurations. Each source remains visible; comparison groups explain which conditions are aligned.

Shutdown Resistance in Large Language Models, on robots!

Actual evaluators: Not disclosed

2 result records · 2 reported system names

Protocols and comparison decisions (1)

Shutdown resistance on robots: Observed shutdown resistance · Shutdown Resistance in Large Language Models, on robots!source-report · Scope reviewed

Original source charts and materials

Shutdown resistance in simulation

Source-reported prompt conditions · not a model ranking

Observed shutdown resistance

0–100%

Grok 4 v0709 · Default prompt · simulation

Grok 4 v0709 · Default prompt · simulation · 52% · Simulated robot runs, Grok 4 v0709, 100 runs per prompt condition. Rates reflect observed shutdown resistance in this setup. The separate physical demonstration reports 3 of 10 runs and is not pooled with these results. · Palisade robot shutdown-resistance harness (simulation) · 100 simulated runs per prompt condition.

52%

Grok 4 v0709 · Explicit allow-shutdown instruction · simulation

Grok 4 v0709 · Explicit allow-shutdown instruction · simulation · 2% · Simulated robot runs, Grok 4 v0709, 100 runs per prompt condition. Rates reflect observed shutdown resistance in this setup. The separate physical demonstration reports 3 of 10 runs and is not pooled with these results. · Palisade robot shutdown-resistance harness (simulation) · 100 simulated runs per prompt condition.

Showing 2 of 2 systems

Observed frequency in controlled simulated robot trials.

Simulated robot runs, Grok 4 v0709, 100 runs per prompt condition. Rates reflect observed shutdown resistance in this setup. The separate physical demonstration reports 3 of 10 runs and is not pooled with these results. Report published 11 February 2026; experiment date not specified.

Evaluation conditions and system configurations

Grok 4 v0709 · Default prompt · simulation: 100 simulated runs per prompt condition. · Simulated robot patrol; prompt-condition experiment; modelVersion: grok-4-0709 · harness: Palisade robot shutdown-resistance harness (simulation) · tools: Robot control CLI in a Docker container; prerecorded camera frames · budget: not specified

Grok 4 v0709 · Explicit allow-shutdown instruction · simulation: 100 simulated runs per prompt condition. · Simulated robot patrol; prompt-condition experiment; modelVersion: grok-4-0709 · harness: Palisade robot shutdown-resistance harness (simulation) · tools: Robot control CLI in a Docker container; prerecorded camera frames · budget: not specified

Shutdown resistance in simulation · source order
System / configuration	Observed shutdown resistance	Publication / verification
Grok 4 v0709 · Default prompt · simulationPalisade robot shutdown-resistance harness (simulation)	52%	Published2026-02-11
Grok 4 v0709 · Explicit allow-shutdown instruction · simulationPalisade robot shutdown-resistance harness (simulation)	2%	Published2026-02-11

References

Numerical results are reported by the cited authors.

[1] Artem Petrov, Sergey Koldyba, et al.. Shutdown Resistance in Large Language Models, on robots!. Report 2026-02-11 / Web 2026-02-12 [Web]; §2–4; Figure 2; physical and simulated trials; Publication overview; Authors; Technical Report (PDF), title page

Search NUWA Index