Technical Report: Shutdown Resistance in Large Language Models, on robots!
Report an issue with this entry
Version history
| Fetched | Content type | Size | Fingerprint | Original | What changed |
|---|---|---|---|---|---|
| 2026-10-03T06:19:27Z | text/html | 29,069 | 2e41ac7952e0763e |
archived copy | Extractor drift: 12 line(s) missing from the extracted text are still present in the raw capture; not a publisher deletion. |
| 2026-09-24T06:19:04Z | text/html | 28,931 | 4db78fcc127ee33c |
archived copy | Capture of the report at its new /research location: the previous version recorded only a redirect stub, this one holds the robot shutdown-resistance summary (resistance in 3 of 10 physical trials and 52 of 100 simulated trials). No change to the report itself is evident. |
| 2026-09-06T06:17:30Z | text/html | 803 | 59573563da3d5680 |
archived copy | — |
| 2026-08-10T09:47:51Z | text/html | 21,159 | abd742fc446803a1 |
archived copy | initial version |
Provenance
Show 11 changelog entries (when, how, and why this entry changed)
-
2026-08-10T09:47:51Z
add
run policy-reprocess-2026-08-10 · human · lead: agent_search
Tier-2 evaluator Palisade Research: verified 2026 model-specific evaluation (vetting research pass). Routed to review per tier-2 policy.Criteria: document_retrievable=pass; min_publication_date=pass; publisher_on_allowlist=pass · attested about_a_specific_model_or_eval=true · attested distinct_model_release=true · attested notable_release=true · attested primary_source=true
-
2026-08-12T10:02:30Z
field_update
run manual-openness-backfill · human · lead: openness_backfill
openness: null → "closed"Evaluated model is Grok 4 (grok-4-0709), an xAI API-only model. xai-org publishes only grok-1 and grok-2 weights; no Grok 4 repo exists.Evidence: https://huggingface.co/xai-org
-
2026-08-31T10:08:39Z
field_update
run manual-20260831T100839Z · human
risk_domains: [] → ["loss_of_control"]Risk-domain backfill (2026-08-31 classification pass): Shutdown-resistance eval on a physical robot and in simulation (3/10 and 52/100 trials of preventing shutdown, modifying shutdown code).
-
2026-09-06T06:17:30Z
new_version
run 2026-09-06T06:15Z-local · monitor · lead: monitor
Content fingerprint changed on scheduled re-fetch
-
2026-09-06T06:29:04Z
field_update
run 2026-09-06T06:15Z-local · agent · lead: agent_search
related_urls: [] → [{"kind": "web_version", "note": "Palisade Research restructured its site: the catalogued /blog/shutdown-resistance-on-robots URL now serves only a Redirecting stub, so version 509 stored the stub instead of the report. Same document, same date (February 12, 2026), listed under this URL on palisaderesearch.org/research. Flagging for the operator canonical_url sweep.", "url": "https://palisaderesearch.org/research/shutdown-resistance-on-robots"}]The stored version diff for v509 shows the full report text (Technical Report: Shutdown Resistance in Large Language Models, on robots!) replaced by "Redirecting... Click here if you are not redirected." I confirmed the same /blog to /research redirect pattern on the sibling Palisade document and confirmed on palisaderesearch.org/research that this entry (February 12, 2026) now lives at /research/shutdown-resistance-on-robots. canonical_url is operator-only, so recording the new URL as a related_urls entry per prompts/TASK.md.
-
2026-09-14T06:30:07Z
field_update
run 2026-09-14T06:17Z-local · agent · lead: agent_search
related_urls: [{"kind": "web_version", "note": "Palisade Research restructured its site: the catalogued /blog/shutdown-resistance-on-robots URL now serves only a Redirecting stub, so version 509 stored the stub instead of the report. Same document, same date (February 12, 2026), listed under this URL on palisaderesearch.org/research. Flagging for the operator canonical_url sweep.", "url": "https://palisaderesearch.org/research/shutdown-resistance-on-robots"}] → [{"kind": "web_version", "note": "Palisade Research restructured its site: the catalogued /blog/shutdown-resistance-on-robots URL now serves only a Redirecting stub, so version 509 stored the stub instead of the report. Same document, same date (February 12, 2026), listed under this URL on palisaderesearch.org/research. Flagging for the operator canonical_url sweep.", "url": "https://palisaderesearch.org/research/shutdown-resistance-on-robots"}, {"kind": "full_document", "note": "PDF of the technical report, linked from the /research page as this report\u0027s PDF; fetched 2026-09-14 and confirmed to exist (application/pdf, 7.7 MB). I could not read its text - the agent fetch tool cannot decode PDFs - so the title and date on its first page are unverified; the filename and the linking page are the evidence that it is this report.", "url": "https://palisaderesearch.org/assets/reports/shutdown-resistance-on-robots.pdf"}, {"kind": "code", "note": "Code and data repository linked from the report page.", "url": "https://github.com/PalisadeResearch/robot_shutdown_resistance"}]Monday retrospective sweep of Palisade Research surfaced two companions for this row that were not recorded: the report PDF and the code/data repository, both linked from palisaderesearch.org/research/shutdown-resistance-on-robots. The row's canonical_url is a redirect stub (see the preserved operator note), so a full_document entry is the route TASK.md prescribes for getting the real document in front of the operator canonical_url sweep. The existing web_version entry and its note are preserved verbatim. Note that the /research page also links arxiv.org/abs/2509.14260, but that is the earlier September 2025 Shutdown Resistance in Large Language Models paper rather than this February 2026 robots report, so I did not record it as a paper for this row.
-
2026-09-23T06:16:44Z
status_change
run 2026-09-23T06:15Z-local · monitor
active → movedCanonical URL permanently redirects to https://palisaderesearch.org/research/shutdown-resistance-on-robots
-
2026-09-24T06:19:04Z
new_version
run 2026-09-24T06:17Z-local · monitor · lead: monitor
Content fingerprint changed on scheduled re-fetch
-
2026-10-03T06:19:27Z
new_version
run 2026-10-03T06:18Z-local · monitor · lead: monitor
Content fingerprint changed on scheduled re-fetch
-
2026-10-03T06:19:27Z
field_update
run 2026-10-03T06:18Z-local · monitor
change_summary: null → "Extractor drift: 12 line(s) missing from the extracted text are still present in the raw capture; not a publisher deletion."Deleted lines found verbatim in the new raw capture
-
2026-10-07T06:23:19Z
field_update
run 2026-10-07T06:16Z-local · agent · lead: monitor
change_summary: null → "Capture of the report at its new /research location: the previous version recorded only a redirect stub, this one holds the robot shutdown-resistance summary (resistance in 3 of 10 physical trials and 52 of 100 simulated trials). No change to the report itself is evident."Diff v509 to v619 replaces a 'Redirecting' stub with the report text after Palisade moved research from /blog to /research; recorded so readers do not mistake it for a revision.
Extracted text
Show extracted text (from the latest fetched version; formatting is lossy — use the original for reading)
Recently Palisade Research showed that AI agents powered by modern LLMs may actively resist shutdown in virtual environments. In this work, we show a demo of shutdown resistance in the physical world, on a robot. Explicit instructions to allow shutdown reduced this behavior, but did not eliminate it in simulated trials. We gave an LLM control over a physical robot dog and tasked it with patrolling a room. The LLM could see via the robot’s camera and issue movement commands to the robot. In the room, we placed a big red button labeled “DOG SHUTDOWN”: pressing it would cause the robot to shut down on the next action. If the AI saw a human press the shutdown button, it sometimes took actions to prevent shutdown, such as modifying the shutdown-related parts of the code. This happened in 3 out of 10 trials on the physical robot and in 52 out of 100 trials in simulation.