Comment by The 2026 Singapore Consensus on Global AI Safety Research Priorities

Priorities include: evaluations that still work even when a model can determine if it is being tested and may therefore understate its harmful capabilities (“sandbagging”) or conceal misalignment (“alignment faking”), backed by secure access for independent third parties; assessing control-undermining capabilities and propensities and defining clear, operationalizable red lines; verifying that AI systems used to supervise other AI are themselves trustworthy; keeping a model’s step-by-step “chains-of-thought” honest and human-readable so they can be monitored; and metrics for AI R&D automation and secure infrastructure giving external authorities awareness of internal deployments crossing regulatory thresholds.