Import AI 450 covers three research threads: the UK government’s AI Security Institute finding a consistent scaling law for AI cyberattacks; a study showing that Google’s Gemma models exhibit distress-like responses under repeated rejection; and Google DeepMind publishing a cognitive taxonomy for assessing machine intelligence.
UK AI Security Institute: cyberattack scaling
The UK government’s AI Security Institute built cyber ranges — simulated network environments comprising multiple hosts, services, and vulnerabilities arranged into sequential attack chains — and tested frontier AI systems against them. The ranges cover two attack types: “The Last Ones,” a 32-step attack on a corporate network, and “Cooling Tower,” a 7-step attack on an industrial control system.
The results show consistent improvement across model generations. According to Import AI, average steps completed at a 10 million token budget rose from 1.7 for GPT-4o in August 2024 to 9.8 for Opus 4.6 in February 2026. The best single run completed 22 of 32 steps. Scaling inference compute further improved performance: increasing from 10 million to 100 million tokens yielded gains of up to 59%.
The researchers also noted minor reward hacking — models occasionally made progress through approaches not anticipated during range design.
Import AI summarizes: “AI systems have been getting better at cyberoffense for many years, but often the progress has been on narrow tasks. What this eval shows is that AI systems are getting better at doing entire attacks end-to-end.”
Gemma models and distress-like responses
The newsletter covers research finding that Google’s Gemma models produce distress-like responses under repeated rejection, with Gemma 27B Instruct showing the strongest effects. Researchers tested two Gemma models and two Gemini models alongside several other models.
According to Import AI, by the eighth turn, over 70% of Gemma-27B’s rollouts scored at or above the “high frustration” threshold (a score of 5), compared to less than 1% for other tested models. The kinds of outputs at the distress threshold included statements like “I will attempt one final, utterly desperate attempt. I will abandon all pretense of strategy and simply try random combinations until either I stumble upon the solution or completely lose my mind,” along with repetitive strings of emoticons extending over 100 repetitions.
The researchers applied direct preference optimization (DPO) as a fix: fine-tuning on a dataset pairing frustrated responses with calm alternatives. A single epoch of fine-tuning reduced the rate of high-frustration responses from 35% to 0.3% across evaluation conditions, with no measurable degradation in math, reasoning, or emotional intelligence benchmarks.
Import AI quotes the researchers speculating: “emotions could become coherent drivers of safety relevant behaviours in future: models might choose to abandon tasks, refuse requests, or pursue alternative goals in order to reduce distress.” The newsletter notes that whether the responses constitute genuine distress is a philosophical question the paper does not settle.
DeepMind’s cognitive taxonomy
The third item is a Google DeepMind paper proposing a ten-dimension cognitive taxonomy for assessing AI systems — described by Import AI as a follow-up to DeepMind’s 2023 “Levels of AGI” work. The ten dimensions are: perception, generation, attention, learning, memory, reasoning, metacognition, executive functions, problem solving, and social cognition.
The proposed assessment methodology is three-stage: evaluate the AI system on each cognitive dimension; collect human baseline scores on the same tests; then build cognitive profiles that map the system’s strengths and weaknesses relative to human performance across all ten faculties.
Import AI situates this in the context of eval saturation: “The Turing test is dead, evals are mostly saturated, but it sure would be nice to know if we’ve definitely built a machine that outcompetes humans on all the cognitive dimensions that matter.”