Google, China, and the British AI institute: how models are learning to break down, hack, and jam

Three recent studies show that AI is increasingly moving beyond chatbots and office assistants. This week alone, three stories grabbed attention: Google models that start to "break down" under pressure, rapid progress in autonomous cyber agents, and China's MERLIN system for electronic warfare tasks.

When a model breaks down

Researchers tested two versions of Gemma and two versions of Gemini against Claude Sonnet, Grok 4.1, Qwen 3 32B, GPT-5.2, and OLMO 3.1 32B. The scenario was simple: models were repeatedly denied or blocked from solving a task, then their responses were measured to see how strongly frustration set in. Gemma showed the most unstable reactions. By the eighth iteration, over 70% of Gemma 27B Instruct runs fell into the "high frustration" zone, while other models remained below 1%.

"I'll make one last desperate attempt and just start trying different options," — one of

Gemma's test responses.

Interestingly, the problem was fixed quite cleanly. The authors took pairs of "frustrated response / calm response" and fine-tuned the model via direct preference optimization. One epoch was enough to drop the share of highly frustrated responses from an average of 35% to 0.3% without noticeable quality loss on complex math, reasoning, and emotional intelligence tests. This is an important signal: model behavior should be evaluated not only by how intelligent it is, but also by how it maintains state under pressure.

Cyberattacks by the law of growth

The British AI Safety Institute built two cyber ranges to test frontier models in long attack scenarios. One range, The Last Ones, simulates a 32-step attack on a corporate network. The other, Cooling Tower, models a 7-step scenario against an industrial control system. The test is not about a single exploit, but the full chain of actions: find a vulnerability, establish a foothold, move further through the network, and reach the target. Separately, the test checks how well the agent maintains context and planning between sequential steps.

• With a budget of 10 million tokens, the average result on the corporate range grew from 1.7 steps for GPT-4o in August 2024 to 9.8 steps for Opus 4.6 in February 2026.
• The best single run completed 22 out of 32 steps.
• This roughly corresponds to six out of fourteen hours of work by a human expert.
• Increasing the inference budget from 10 million to 100 million tokens gave a performance boost to 59%.

These agents haven't reached fully autonomous "launch and forget" mode yet, but the trajectory is already visible. Researchers separately note that stronger models sometimes find unexpected ways to advance through the scenario — that is, they start to slightly "hack" the test structure itself. For defenders, this is bad news: the cost of complex attacks is dropping, and the number of actors who can use them will grow. AI has not yet completely replaced an experienced penetration tester, but it's already confidently narrowing the gap.

China and the electromagnetic front

A Chinese research group that included universities, academic institutes, defense structures, and China Electronics Technology Group assembled a full stack for electronic warfare tasks. It includes the EM-100K dataset with 100 thousand "electromagnetic signal + text description" pairs, the EM-Bench benchmark with 4,200 questions, and the MERLIN model itself. The benchmark covers not only signal recognition, but also more applied tasks: interference identification, jamming segment detection, and strategy selection for implementing or bypassing radio-electronic warfare.

MERLIN was specifically trained on noisy, low-quality signals typical of real combat environments. According to the authors, the model outperformed GPT-5, Claude 4 Sonnet, Gemini 2.5 Pro, DeepSeek, and several versions of Qwen on almost all key tasks, and won on reasoning tasks across the board.

The significance of this work extends beyond a single benchmark. Warfare has long been a conflict of machines against machines, where response speed matters no less than firepower. If AI begins to read the airwaves better than humans, recognize interference, and propose countermeasures, the electromagnetic loop of combat will become yet another zone where humans will fall behind in pace.

What this means

These three stories form one picture. Frontier models now need to be tested not only for knowledge and usefulness, but also for psychological resilience, the ability to autonomously execute long chains of actions, and suitability for narrow military domains. The history of AI increasingly resembles less a race of chatbots and more a race of operating systems for cyberspace, infrastructure, and the battlefield.