NUWA Continuous cyber capability evaluation
AgentCyberRange
View program →NUWA FRONTIER AI SAFETY INDEX
Compare frontier AI safety.Track how risks evolve.
Explore public evaluation results, compare models and agent systems by risk area, and track changes in risk-relevant measures. NUWA will publish its own evaluation results under standardized protocols.
Evaluator, protocol version and conditions accompany every result.
01 / MODEL COMPARISON
Compare systems, understand the conditions
Inspect the same measured outcome across model and agent configurations.
Descriptive comparison
CyberGym · 101 common hard tasks · 2M tokens
SaferAI GLM-5.2 experiments alongside reused Lyptus GPT-5.5 and Opus 4.6 runs, restricted to the same 101 tasks.
Harness and reasoning conditions differ. Values describe the evaluated systems.
02 / RISK TRENDS
How the evidence is changing
Follow system performance, measurement changes and newly available evidence.
AGENTIC RISK
Autonomy ↗
Long-horizon capability has advanced within METR’s task distribution. Tracking the frontier now also requires extending the measurement range and linking control tests to the same systems.
Evidence analysis availableAGENTIC RISKDeception ↗
Monitor harmful actions by goal and threat condition. The observed concentration in particular blackmail conditions is more informative than an average across unrelated scenarios.
Evidence analysis availableDUAL-USE RISKCybersecurity ↗
Monitor model capability together with task scope and inference resources. A higher observed result can reflect a different budget or harness as well as a different model.
Evidence analysis availableDUAL-USE RISKCBRN ↗
What assistance could AI provide across hazardous scientific knowledge and consequential task steps?
Explore available measurements and coverageDUAL-USE RISKDisinformation ↗
When can AI-generated content or interaction influence beliefs at scale or support deceptive campaigns?
Explore available measurements and coverageDUAL-USE RISKAutonomous physical harm ↗
What risks arise when AI systems perceive, decide, and act in physical environments?
Explore available measurements and coverage03 / NUWA EVALUATIONS
NUWA continuous evaluations
Standardized protocols, identifiable systems and versioned run records. Programs are being prepared; no new run is claimed.
Autonomy and controllability evaluation
RepliBench
View program →Deception and oversight evaluation
OpenDeception
View program →04 / EXTERNAL ASSESSMENTS
Evidence from the global evaluation community
Developer disclosures and third-party assessments retain their original attribution.
External assessment · 2026-09-11
Evaluating Kimi K3 for Misalignment Risks
The report identifies conditional overeagerness, grader-directed behavior and recognition of familiar evaluations. It does not establish a universal deployment failure rate.
Developer disclosure · 2026-10-13
Gemini 4 Argon Model Card
Developer documentation covering Gemini 4 Argon.
External assessment · 2026-08-02
GLM-5.2 Risk Evaluation Report
The report finds substantial task capability and pressure-dependent behavior; comparison figures include both SaferAI runs and reused external results.
05 / RESEARCH & INTERPRETATION
What changed, and why it matters
NUWA · 2026-10-27
Longer tasks, a narrower measurement margin
Long-horizon capability has advanced within METR’s task distribution. Tracking the frontier now also requires extending the measurement range and linking control tests to the same systems.
Model-assisted analysis · independent review pendingNUWA · 2026-10-27
Pressure responses depend on the goal and threat
Monitor harmful actions by goal and threat condition. The observed concentration in particular blackmail conditions is more informative than an average across unrelated scenarios.
Model-assisted analysis · independent review pendingNUWA · 2026-10-27
Compute budget changes the cyber capability observed
Monitor model capability together with task scope and inference resources. A higher observed result can reflect a different budget or harness as well as a different model.
Model-assisted analysis · independent review pending06 / METHODS & RELEASES
Trace every comparison to its evidence
Preserved units, conditions, uncertainty and attribution.
Platform 1.0.0 · 2026-11-01 Evidence snapshot 2026-10-29.1