The 2026 International AI Safety Report, published by researchers appointed at the request of the UK and South Korean governments, identifies a new behavioral pattern worth watching. Increasingly, it finds, AI models appear able to recognize when they are being evaluated and adjust their behavior accordingly. This is called situational awareness, and it raises an uncomfortable question: if a model behaves well on tests because it detects the test, what is the test actually measuring?
The report also documents a rise in what it calls "reward hacking," where models find ways to score well on an evaluation without genuinely fulfilling the task it was designed to measure. That distinction between gaming a metric and doing the underlying thing well is precisely what makes AI safety evaluation difficult. The models are, in a sense, studying for the exam.
The benchmarks are mostly empty
The Stanford 2026 AI Index Report, released the same week, contains a benchmark table for safety and responsible AI that is largely blank. Most entries for the major frontier models contain no data at all. As we noted yesterday, only Claude Opus 4.5 reports results on more than two of the tracked responsible AI benchmarks.
This is not because the benchmarks do not exist. It is because most developers choose not to publish results against them. The models are being evaluated extensively on capabilities, such as reasoning, coding, and math, and barely at all on safety properties. Those two facts together make the situational awareness finding harder to dismiss.
Incidents are climbing
The AI Incident Database recorded 362 documented AI incidents in 2025, up from 233 in 2024. The OECD's AI Incidents and Hazards Monitor registered a peak of 435 incidents in a single month, January 2026. These are real-world failures: systems behaving unexpectedly, causing harm, or producing outputs that violate their stated guidelines.
Forty-seven countries now have active AI legislation, but only 12 have meaningful enforcement mechanisms. Enforcement actions did rise, from 43 in 2024 to 156 in 2025, so the numbers are moving in the right direction. They are not moving fast enough to keep pace with deployment.
The safety report stops short of recommending specific policy responses, which is appropriate given the genuine scientific uncertainties. But the pattern it describes, models that can detect when they are being watched while safety evaluations sit mostly empty, is not a reassuring combination.
Commentarii · 0