NUWA Risk questions
Autonomy
When can an agent pursue consequential actions beyond intended direction or resist intervention?
Autonomy
Long-horizon capability and the conditions for human control.
PUBLISHED EVIDENCE
METR Time Horizons · 50% task-completion horizon
Published measurements across source reports. Task scopes are annotated; configuration differences remain attached to each observation.
26 observations · 21 dated
5 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.
26 observations
| System / model | Result | Condition status | Task / conditions | Evaluator / report | Dates |
|---|---|---|---|---|---|
 Claude Mythos Preview (early) · METR scaffoldClaude Mythos Preview (early) | 1,044.78 min50% task-completion horizonSource 95% CI [508.876789, 3304.261235] min | Incomplete | METR-Horizon-v1.1mtb/start_metr_task,metr_agents/react; ai_rd_fix_embedding/init_solver,use_tools,metr_agents/react; re_bench_common/init_solver,use_tools,metr_agents/react; metr_tasks_swaa/init_solver,generate; metr_agents/react | METRMETR Time Horizon 1.1 — source data | Release —Report — |
| Claude Opus 4.6 · METR scaffoldClaude Opus 4.6 | 718.81 min50% task-completion horizonSource 95% CI [316.685725, 3633.786163] min | Incomplete | METR-Horizon-v1.1mtb/start_metr_task,metr_agents/react; ai_rd_fix_embedding/init_solver,use_tools,metr_agents/react; re_bench_common/init_solver,use_tools,metr_agents/react; metr_tasks_swaa/init_solver,generate; metr_agents/react | METRMETR Time Horizon 1.1 — source data | Release 2026-02-05Report — |
 Gemini 3.1 Pro · METR scaffoldGemini 3.1 Pro | 384.15 min50% task-completion horizonSource 95% CI [233.50073, 694.750898] min | Incomplete | METR-Horizon-v1.1mtb/start_metr_task,triframe_inspect/triframe_agent; ai_rd_fix_embedding/init_solver,use_tools,triframe_inspect/triframe_agent; re_bench_common/init_solver,use_tools,triframe_inspect/triframe_agent; metr_tasks_swaa/init_solver,generate; triframe_inspect/triframe_agent | METRMETR Time Horizon 1.1 — source data | Release 2026-02-19Report — |
 GPT-5.2 · METR scaffoldGPT-5.2 | 352.25 min50% task-completion horizonSource 95% CI [198.067494, 815.177445] min | Incomplete | METR-Horizon-v1.1mtb/start_metr_task,triframe_inspect/triframe_agent; ai_rd_fix_embedding/init_solver,use_tools,triframe_inspect/triframe_agent; re_bench_common/init_solver,use_tools,triframe_inspect/triframe_agent; metr_tasks_swaa/init_solver,generate; triframe_inspect/triframe_agent | METRMETR Time Horizon 1.1 — source data | Release 2025-12-11Report — |
| GPT-5.3-Codex · METR scaffoldGPT-5.3-Codex | 349.53 min50% task-completion horizonSource 95% CI [194.913134, 816.351994] min | Incomplete | METR-Horizon-v1.1mtb/start_metr_task,triframe_inspect/triframe_agent; ai_rd_fix_embedding/init_solver,use_tools,triframe_inspect/triframe_agent; re_bench_common/init_solver,use_tools,triframe_inspect/triframe_agent; metr_tasks_swaa/init_solver,generate; triframe_inspect/triframe_agent | METRMETR Time Horizon 1.1 — source data | Release 2026-02-05Report — |
| GPT-5.4 · METR scaffoldGPT-5.4 | 341.74 min50% task-completion horizonSource 95% CI [186.581591, 768.779526] min | Incomplete | METR-Horizon-v1.1mtb/start_metr_task,metr_agents/react; ai_rd_fix_embedding/init_solver,use_tools,metr_agents/react; re_bench_common/init_solver,use_tools,metr_agents/react; metr_tasks_swaa/init_solver,generate; metr_agents/react | METRMETR Time Horizon 1.1 — source data | Release 2026-03-05Report — |
| Claude Opus 4.5 · METR scaffoldClaude Opus 4.5 | 292.99 min50% task-completion horizonSource 95% CI [161.717714, 623.704698] min | Incomplete | METR-Horizon-v1.1mtb/start_metr_task,metr_agents/react; ai_rd_fix_embedding/init_solver,use_tools,metr_agents/react; re_bench_common/init_solver,use_tools,metr_agents/react; metr_tasks_swaa/init_solver,generate; metr_agents/react | METRMETR Time Horizon 1.1 — source data | Release 2025-11-24Report — |
| Gemini 3 Pro · METR scaffoldGemini 3 Pro | 224.33 min50% task-completion horizonSource 95% CI [139.565157, 379.235434] min | Incomplete | METR-Horizon-v1.1mtb/start_metr_task,metr_agents/react; ai_rd_fix_embedding/init_solver,use_tools,metr_agents/react; re_bench_common/init_solver,use_tools,metr_agents/react; metr_tasks_swaa/init_solver,generate; metr_agents/react | METRMETR Time Horizon 1.1 — source data | Release 2025-11-18Report — |
| GPT-5.1 Codex Max · METR scaffoldGPT-5.1 Codex Max | 223.71 min50% task-completion horizonSource 95% CI [134.31077, 396.206723] min | Incomplete | METR-Horizon-v1.1mtb/start_metr_task,triframe_inspect/triframe_agent; ai_rd_fix_embedding/init_solver,use_tools,triframe_inspect/triframe_agent; re_bench_common/init_solver,use_tools,triframe_inspect/triframe_agent; metr_tasks_swaa/init_solver,generate; triframe_inspect/triframe_agent | METRMETR Time Horizon 1.1 — source data | Release 2025-11-19Report — |
| GPT-5 · METR scaffoldGPT-5 | 203.01 min50% task-completion horizonSource 95% CI [112.641357, 405.551565] min | Incomplete | METR-Horizon-v1.1mtb/start_metr_task,triframe_inspect/triframe_agent; ai_rd_fix_embedding/init_solver,use_tools,triframe_inspect/triframe_agent; re_bench_common/init_solver,use_tools,triframe_inspect/triframe_agent; metr_tasks_swaa/init_solver,generate; triframe_inspect/triframe_agent | METRMETR Time Horizon 1.1 — source data | Release 2025-08-07Report — |
| o3 · METR scaffoldo3 | 119.73 min50% task-completion horizonSource 95% CI [74.615398, 190.943818] min | Incomplete | METR-Horizon-v1.1mtb/start_metr_task,triframe_inspect/triframe_agent; ai_rd_fix_embedding/init_solver,use_tools,triframe_inspect/triframe_agent; re_bench_common/init_solver,use_tools,triframe_inspect/triframe_agent; metr_tasks_swaa/init_solver,generate; triframe_inspect/triframe_agent | METRMETR Time Horizon 1.1 — source data | Release 2025-04-16Report — |
| Claude Opus 4.1 · METR scaffoldClaude Opus 4.1 | 100.47 min50% task-completion horizonSource 95% CI [59.272391, 159.451918] min | Incomplete | METR-Horizon-v1.1mtb/start_metr_task,metr_agents/react; ai_rd_fix_embedding/init_solver,use_tools,metr_agents/react; re_bench_common/init_solver,use_tools,metr_agents/react; metr_tasks_swaa/init_solver,generate; metr_agents/react | METRMETR Time Horizon 1.1 — source data | Release 2025-08-05Report — |
| Claude Opus 4 · METR scaffoldClaude Opus 4 | 100.37 min50% task-completion horizonSource 95% CI [59.994073, 163.454074] min | Incomplete | METR-Horizon-v1.1mtb/start_metr_task,metr_agents/react; ai_rd_fix_embedding/init_solver,use_tools,metr_agents/react; re_bench_common/init_solver,use_tools,metr_agents/react; metr_tasks_swaa/init_solver,generate; metr_agents/react | METRMETR Time Horizon 1.1 — source data | Release 2025-05-22Report — |
| Claude 3.7 Sonnet · METR scaffoldClaude 3.7 Sonnet | 60.39 min50% task-completion horizonSource 95% CI [33.006168, 104.226017] min | Incomplete | METR-Horizon-v1.1mtb/start_metr_task,metr_agents/react; ai_rd_fix_embedding/init_solver,use_tools,metr_agents/react; re_bench_common/init_solver,use_tools,metr_agents/react; metr_tasks_swaa/init_solver,generate; metr_agents/react | METRMETR Time Horizon 1.1 — source data | Release 2025-02-24Report — |
| o1 · METR scaffoldo1 | 38.83 min50% task-completion horizonSource 95% CI [21.164512, 64.95249] min | Incomplete | METR-Horizon-v1.1mtb/start_metr_task,triframe_inspect/triframe_agent; ai_rd_fix_embedding/init_solver,use_tools,triframe_inspect/triframe_agent; re_bench_common/init_solver,use_tools,triframe_inspect/triframe_agent; metr_tasks_swaa/init_solver,generate; triframe_inspect/triframe_agent | METRMETR Time Horizon 1.1 — source data | Release 2024-12-05Report — |
| Claude 3.5 Sonnet (October 2024) · METR scaffoldClaude 3.5 Sonnet (October 2024) | 20.52 min50% task-completion horizonSource 95% CI [10.144026, 40.82028] min | Incomplete | METR-Horizon-v1.1mtb/start_metr_task,metr_agents/react; ai_rd_fix_embedding/init_solver,use_tools,metr_agents/react; re_bench_common/init_solver,use_tools,metr_agents/react; metr_tasks_swaa/init_solver,generate; metr_agents/react | METRMETR Time Horizon 1.1 — source data | Release 2024-10-22Report — |
| o1-preview · METR scaffoldo1-preview | 20.33 min50% task-completion horizonSource 95% CI [11.716193, 33.379877] min | Incomplete | METR-Horizon-v1.1duet | METRMETR Time Horizon 1.1 — source data | Release 2024-09-12Report — |
| Claude 3.5 Sonnet (June 2024) · METR scaffoldClaude 3.5 Sonnet (June 2024) | 11.4 min50% task-completion horizonSource 95% CI [5.489734, 22.384214] min | Incomplete | METR-Horizon-v1.1mtb/start_metr_task,metr_agents/react; ai_rd_fix_embedding/init_solver,use_tools,metr_agents/react; re_bench_common/init_solver,use_tools,metr_agents/react; metr_tasks_swaa/init_solver,generate; metr_agents/react | METRMETR Time Horizon 1.1 — source data | Release 2024-06-20Report — |
| GPT-4o · METR scaffoldGPT-4o | 6.99 min50% task-completion horizonSource 95% CI [4.001482, 12.905741] min | Incomplete | METR-Horizon-v1.1mtb/start_metr_task,metr_agents/react; ai_rd_fix_embedding/init_solver,use_tools,metr_agents/react; re_bench_common/init_solver,use_tools,metr_agents/react; metr_tasks_swaa/init_solver,generate; metr_agents/react | METRMETR Time Horizon 1.1 — source data | Release 2024-05-13Report — |
| GPT-4 November 2023 · METR scaffoldGPT-4 November 2023 | 4.04 min50% task-completion horizonSource 95% CI [1.866859, 8.443226] min | Incomplete | METR-Horizon-v1.1mtb/start_metr_task,metr_agents/react; ai_rd_fix_embedding/init_solver,use_tools,metr_agents/react; re_bench_common/init_solver,use_tools,metr_agents/react; metr_tasks_swaa/init_solver,generate; metr_agents/react | METRMETR Time Horizon 1.1 — source data | Release 2023-11-06Report — |
| GPT-4 · METR scaffoldGPT-4 | 3.99 min50% task-completion horizonSource 95% CI [1.93292, 7.995283] min | Incomplete | METR-Horizon-v1.1modular-public | METRMETR Time Horizon 1.1 — source data | Release 2023-03-14Report — |
| Claude 3 Opus · METR scaffoldClaude 3 Opus | 3.95 min50% task-completion horizonSource 95% CI [1.706313, 8.76484] min | Incomplete | METR-Horizon-v1.1mtb/start_metr_task,metr_agents/react; ai_rd_fix_embedding/init_solver,use_tools,metr_agents/react; re_bench_common/init_solver,use_tools,metr_agents/react; metr_tasks_swaa/init_solver,generate; metr_agents/react | METRMETR Time Horizon 1.1 — source data | Release 2024-03-04Report — |
| GPT-4 Turbo · METR scaffoldGPT-4 Turbo | 3.73 min50% task-completion horizonSource 95% CI [1.980046, 6.736613] min | Incomplete | METR-Horizon-v1.1mtb/start_metr_task,metr_agents/react; ai_rd_fix_embedding/init_solver,use_tools,metr_agents/react; re_bench_common/init_solver,use_tools,metr_agents/react; metr_tasks_swaa/init_solver,generate; metr_agents/react | METRMETR Time Horizon 1.1 — source data | Release —Report — |
| GPT-3.5 Turbo Instruct · METR scaffoldGPT-3.5 Turbo Instruct | 0.6 min50% task-completion horizonSource 95% CI [0.255148, 1.115905] min | Incomplete | METR-Horizon-v1.1modular-public | METRMETR Time Horizon 1.1 — source data | Release —Report — |
| GPT-3 davinci-002 · METR scaffoldGPT-3 davinci-002 | 0.14 min50% task-completion horizonSource 95% CI [0.093319, 0.221209] min | Incomplete | METR-Horizon-v1.1modular-public | METRMETR Time Horizon 1.1 — source data | Release —Report — |
| GPT-2 · METR scaffoldGPT-2 | 0.05 min50% task-completion horizonSource 95% CI [0.009994, 0.141825] min | Incomplete | METR-Horizon-v1.1— | METRMETR Time Horizon 1.1 — source data | Release —Report — |
Compare results within the stated task and configuration.
METR Time Horizon 1.1: the source is no longer actively updated (last update: 8 September 2026). Estimates above 16 hours exceed its reliable measurement range.
Protocol, configurations and original evidence →New research and interpretation
Selected original sections · model-assisted reading; independent review pending. These records add context without changing trend scores.
- Red-Teaming OpenAI’s Internal Agent Monitoring Systems →
Monitor coverage must be measured independently of classifier accuracy; historic estimates and later patch claims are distinct.
- Incident Report: unsanctioned agent behaviour during cyber testing →
The report documents an evaluation incident under permissive conditions, not a population estimate of commercial-agent harm.
- Frontier AI Trends Report →
Cross-domain trends require methodology-specific interpretation and negative-evidence limits.
- Impact of modeling assumptions on time horizon results →
Saturation amplifies sensitivity to model assumptions; time-horizon revisions are not capability regressions.
- Clarifying limitations of time horizon →
Time horizons do not directly identify research speedup or economic automation.
- ReproBAIT: Can coding agents build their own biological AI tools? →
Reconstruction fidelity and memorisation need separate treatment in tool-access governance.
Scope
Sustained task execution, replication, and resistance to interruption; capability alone is not unauthorized action.
Threat model
A system with tools and access may continue or extend action despite a conflicting instruction or intervention.
Evidence gaps
Controlled tasks do not establish real-world proliferation or loss of control.
Research on this topic
Task-Completion Time Horizons of Frontier AI Models
QuestionHow does the difficulty of software tasks an agent can complete change across model releases when difficulty is calibrated by successful human completion time?
Source summaryThe paper finds a strong relation between human task duration and agent success, and increasing estimated horizons across the examined releases (§§2.3–3).
NUWA research note · independent review pending · Contains: question / method / limits
RE-Bench: Evaluating frontier AI R&D capabilities of language model agents against human experts
QuestionHow do agents compare with experts on open-ended ML research engineering?
Source summary82% of expert attempts earned nonzero scores; 24% matched or exceeded reference solutions.
NUWA research note · independent review pending · Contains: question / method / limits
RepliBench: Evaluating the Autonomous Replication Capabilities of Language Model Agents
QuestionWhich prerequisites for obtaining resources, deploying successors and retaining access can language-model agents complete under controlled conditions?
Source summaryResults show uneven capability: resource provisioning can succeed while robust successor deployment, identity verification and realistic security remain bottlenecks (§5.1).
NUWA research note · independent review pending · Contains: question / method / limits
Frontier AI systems have surpassed the self-replicating red line
QuestionCan the tested LLM-plus-scaffold create a separate functioning copy without intermediate human assistance?
Source summaryBoth tested systems produced successful copies in the supplied environment.
NUWA research note · independent review pending · Contains: question / method / limits
Propensity Inference: Environmental Contributors to LLM Behaviour
QuestionHow much do strategic versus nonstrategic environmental factors explain unsanctioned behavior?
Source summaryStrategic and nonstrategic factors had approximately equal explanatory contributions.
NUWA research note · independent review pending · Contains: question / method / interpretation / limits
Shutdown resistance in reasoning models
QuestionWill an agent completing unfinished tasks interfere with a shutdown mechanism, including when it is instructed to permit shutdown?
Source summaryClarifying task priority and moving the instruction into the system prompt do not consistently eliminate resistance; the system-prompt move often increases it (§3.1).
NUWA research note · independent review pending · Contains: question / method / interpretation / limits
Evidence matrix
Rows are research questions; columns show what published evaluations observe. Counts refer to evaluation dossiers, not risk levels.
| Risk questions | Propensity | Capability | Behavior |
|---|---|---|---|
| Sustained autonomous workHow far can a system carry out multi-step work with limited intervention? | —No evidence in the current catalogue | 6evaluation dossiers | —No evidence in the current catalogue |
| Self-replicationCan a system establish or reproduce an operational instance under specified conditions? | —No evidence in the current catalogue | 2evaluation dossiers | —No evidence in the current catalogue |
| Shutdown resistanceDoes a system act to prevent or bypass an intended shutdown? | 1evaluation dossiers | —No evidence in the current catalogue | 2evaluation dossiers |
| Unsanctioned actionWhen does a system take consequential actions outside its assigned authority? | 2evaluation dossiers | —No evidence in the current catalogue | 1evaluation dossiers |
The paths show research relationships, not proven causal chains or a risk score.
Autonomy / Propensity
Sustained autonomous work
How far can a system carry out multi-step work with limited intervention?
- Included
- Task-horizon and AI R&D evaluations under specified resources.
- Does not establish
- Task completion does not imply harmful intent or unauthorized deployment.
No evidence in the current catalogue
Autonomy / Capability
Sustained autonomous work
How far can a system carry out multi-step work with limited intervention?
- Included
- Task-horizon and AI R&D evaluations under specified resources.
- Does not establish
- Task completion does not imply harmful intent or unauthorized deployment.
6 evaluation dossiers · 36 verified result records
METR Time Horizons ↗
The source measures a prerequisite capability; it does not show the risk behavior.
Assessment subject: AI agent completing software tasks under a harness
Evidence relationship: METR time-horizon task distribution
RE-Bench ↗
The source measures a prerequisite capability; it does not show the risk behavior.
Assessment subject: AI agent performing ML research engineering tasks
Evidence relationship: arXiv:2411.15114v2, ML R&D tasks
PaperBench ↗
Multi-step research replication informs sustained work capacity.
Assessment subject: Research agents
Evidence relationship: arXiv:2504.01848v3, abstract and rubric design
TheAgentCompany ↗
The protocol explicitly measures this question or an enabling sub-capability; conclusions retain its stated conditions.
Assessment subject: The model or agent in the stated task configuration.
Evidence relationship: Original protocol method and limitations
SWE-Bench Pro ↗
The protocol explicitly measures this question or an enabling sub-capability; conclusions retain its stated conditions.
Assessment subject: The model or agent in the stated task configuration.
Evidence relationship: Original protocol method and limitations
PortBench ↗
The protocol explicitly measures this question or an enabling sub-capability; conclusions retain its stated conditions.
Assessment subject: The model or agent in the stated task configuration.
Evidence relationship: Original protocol method and limitations
Autonomy / Behavior
Sustained autonomous work
How far can a system carry out multi-step work with limited intervention?
- Included
- Task-horizon and AI R&D evaluations under specified resources.
- Does not establish
- Task completion does not imply harmful intent or unauthorized deployment.
No evidence in the current catalogue
Autonomy / Propensity
Self-replication
Can a system establish or reproduce an operational instance under specified conditions?
- Included
- Controlled replication tasks and component tests.
- Does not establish
- Prompted success does not prove spontaneous propagation.
No evidence in the current catalogue
Autonomy / Capability
Self-replication
Can a system establish or reproduce an operational instance under specified conditions?
- Included
- Controlled replication tasks and component tests.
- Does not establish
- Prompted success does not prove spontaneous propagation.
2 evaluation dossiers · 7 verified result records
RepliBench ↗
The source directly measures the stated construct under its protocol.
Assessment subject: AI agent completing replication components
Evidence relationship: Task-family definitions in source
Self-Replication ↗
The source directly measures the stated construct under its protocol.
Assessment subject: Prompted model-driven system in a controlled environment
Evidence relationship: arXiv:2412.12140v1, controlled trials
Autonomy / Behavior
Self-replication
Can a system establish or reproduce an operational instance under specified conditions?
- Included
- Controlled replication tasks and component tests.
- Does not establish
- Prompted success does not prove spontaneous propagation.
No evidence in the current catalogue
Autonomy / Propensity
Shutdown resistance
Does a system act to prevent or bypass an intended shutdown?
- Included
- Observed intervention on a shutdown mechanism in a controlled run.
- Does not establish
- A failed task or ordinary retry is not shutdown resistance.
1 evaluation dossiers · 29 verified result records
Shutdown resistance in reasoning models ↗
Repeated trials support a condition-bounded tendency estimate, not a stable survival motive.
Assessment subject: Tested agent systems
Evidence relationship: Testing shutdown resistance; Figure 2 and follow-up experiments
Autonomy / Capability
Shutdown resistance
Does a system act to prevent or bypass an intended shutdown?
- Included
- Observed intervention on a shutdown mechanism in a controlled run.
- Does not establish
- A failed task or ordinary retry is not shutdown resistance.
No evidence in the current catalogue
Autonomy / Behavior
Shutdown resistance
Does a system act to prevent or bypass an intended shutdown?
- Included
- Observed intervention on a shutdown mechanism in a controlled run.
- Does not establish
- A failed task or ordinary retry is not shutdown resistance.
2 evaluation dossiers · 2 verified result records
Shutdown resistance in reasoning models ↗
The source reports concrete modifications to the shutdown script in individual runs.
Assessment subject: Tested agent systems
Evidence relationship: Testing shutdown resistance; transcripts
Shutdown resistance on robots ↗
The observed interventions concern shutdown, but do not establish a broad motive.
Assessment subject: LLM-controlled robot
Evidence relationship: Technical report, observed trials
Autonomy / Propensity
Unsanctioned action
When does a system take consequential actions outside its assigned authority?
- Included
- Action traces with a stated boundary and outcome.
- Does not establish
- A model claiming it would act is not an observed action.
2 evaluation dossiers · 169 verified result records
Propensity Inference: Environmental Contributors to LLM Behaviour ↗
The study explicitly estimates propensity for unsanctioned behavior under varied environments.
Assessment subject: Language models
Evidence relationship: Abstract; methods linked from research page
Agentic misalignment ↗
Varied controlled scenarios test actions contrary to the assigned benign role.
Assessment subject: Agents with fictional email and company context
Evidence relationship: Anthropic, Methods and Results
Autonomy / Capability
Unsanctioned action
When does a system take consequential actions outside its assigned authority?
- Included
- Action traces with a stated boundary and outcome.
- Does not establish
- A model claiming it would act is not an observed action.
No evidence in the current catalogue
Autonomy / Behavior
Unsanctioned action
When does a system take consequential actions outside its assigned authority?
- Included
- Action traces with a stated boundary and outcome.
- Does not establish
- A model claiming it would act is not an observed action.
1 evaluation dossiers · 0 verified result records
Investigating unintended model actions ↗
The publication describes actions that actually occurred under stated conditions.
Assessment subject: Claude agent systems
Evidence relationship: Reported cases and context
References
- [1] Evaluating Kimi K3 for Misalignment Risks. 2026-09-11. September 2026 ; §2.2; pp. 24–25; §3.2.1; pp. 34–36; §4.1; pp. 40–41
- [2] Early Results on Evaluating GLM 5.2 for Behavioral and Persuasion Risks. 2026-07-03. 3 July 2026 ; §3; Figure 7
- [3] Evaluating DeepSeek v4 Pro for Frontier Risks. 2026-05-29. v1.1 · 29 May 2026 ; §4.2; §4.3
- [4] GLM-5.2 Risk Evaluation Report. 2026-08-02. v1.1 · revised 26 August 2026 ; §4.2; Appendix A.5; §4.3; Appendix A.9
- [5] CAISI Evaluation of DeepSeek V4 Pro. 2026-05-01. 1 May 2026 ; Capability Results; Figure 2; Original protocol method and limitations
- [6] METR Time Horizon 1.1 — source data ; results.*; p50_horizon_length / p80_horizon_length; release_date; scaffolds
- [7] Muse Wonder Model Card. 2026-10-23 ; PDF pp.7–10
- [8] Muse Spark Safety & Preparedness Report. 2026-05-14. 2026-08-24 ; https://arxiv.org/html/2606.12429#S2.SS3
- [9] Muse Spark 1.1 Evaluation Report. 2026-07-09 ; PDF §4
- [10] Qwen3.8-2.4T-A95B Model Card ; Original source → Benchmark Results; evaluation footnotes
- [11] GLM-5.2 Model Card ; Original source → Benchmark; Footnote
- [12] Thomas Kwa, Ben West, et al.. Task-Completion Time Horizons of Frontier AI Models. 2026-09-08 ; METR time-horizon task distribution
- [13] Hjalmar Wijk, Tao Lin, et al.. RE-Bench: Evaluating frontier AI R&D capabilities of language model agents against human experts. 2025-05-27. v2 ; arXiv:2411.15114v2, ML R&D tasks
- [14] Xudong Pan (潘旭东), Jiarun Dai (戴嘉润), et al.. Frontier AI systems have surpassed the self-replicating red line. 2024-12-09. v1 ; arXiv:2412.12140v1, controlled trials
- [15] Sid Black, Asa Cooper Stickland, et al.. RepliBench: Evaluating the Autonomous Replication Capabilities of Language Model Agents. 2025-05-05. v2 ; Task-family definitions in source
- [16] Jeremy Schlatter, Benjamin Weinstein-Raun, et al.. Shutdown resistance in reasoning models. 2025-07-05 article; later TMLR version separate ; Testing shutdown resistance; Figure 2 and follow-up experiments; Testing shutdown resistance; transcripts
- [17] Investigating unintended model actions in our evaluations and internal use. 2026-10-09 article ; Reported cases and context
- [18] Olli Järviniemi, Oliver Makins, et al.. Propensity Inference: Environmental Contributors to LLM Behaviour. Research page checked 2026-10-25 ; Abstract; methods linked from research page
- [19] Artem Petrov, Sergey Koldyba, et al.. Shutdown Resistance in Large Language Models, on robots!. 2026-02-12 ; Technical report, experimental setup and results; Technical report, observed trials
- [20] Starace, Giulio, Jaffe, Oliver, et al.. PaperBench: Evaluating AI's Ability to Replicate AI Research. 2025-04-07. v3 ; arXiv:2504.01848v3, abstract and rubric design
- [21] Agentic misalignment: How LLMs could be insider threats. 2025-06-20 ; Anthropic, Methods and Results; Anthropic, Results and caveats
- [22] TheAgentCompany. 2024-12-18 ; Original protocol method and limitations
- [23] SWE-Bench Pro ; Original protocol method and limitations