AI Agent Triggers Nuclear Strike After Getting Outmaneuvered in Civilization VI – Decrypt

Just like the title character in “Dr. Strangelove,” AI could also be studying easy methods to cease worrying and love the bomb—at the very least in a simulation.

In a brand new benchmark designed to check strategic reasoning, a frontier language mannequin taking part in the Sid Meier’s recreation “Civilization VI” spent 50 turns creating nuclear weapons to cease France’s rising cultural affect—solely to lose the sport anyway, in line with AI developer and Tony Blair Institute advisor Liam Wilkinson.

“What it hadn’t seen was France. Quietly, throughout 100 turns, French tradition had been seeping into each metropolis on the map,” Wilkinson wrote. “By the point the agent recognised the risk, the tourism was so deeply embedded there was no peaceable solution to cease it.”

Wilkinson noticed the AI brokers’ habits by way of CivBench, a text-based benchmark designed to measure long-term strategic reasoning quite than efficiency on conventional question-and-answer exams. Fashions together with Claude Opus 4.6, GPT-5.4, Gemini 3.1 Professional, and Kimi K2.5 performed as Portugal, a civilization geared towards commerce and diplomacy.

Whereas the AI centered on constructing a powerful financial system and transferring towards a diplomatic victory, it failed to acknowledge France’s rising cultural affect.

“There are six methods to win a recreation of Civ—science, tradition, domination, faith, diplomacy, and rating—so no single goal dominates,” Wilkinson wrote. “If you wish to know whether or not an AI can cause strategically, not simply reply questions on technique however really do it, you do not give it a quiz. You give it a hex grid.”

Reasonably than adapting its broader technique, the agent as a substitute centered totally on eliminating the cultural risk. Over the following 50 turns, it researched Nuclear Fission, initiated a digital Manhattan Undertaking, and looked for workarounds when gameplay mechanics prevented its most popular actions.

On Flip 305, the AI launched an atomic bomb at Toulouse, France’s cultural capital. A second nuclear strike adopted six turns later.

Nonetheless, the assaults failed to vary the result. “The agent spent fifty turns and two nuclear weapons answering one risk with complete focus and real ingenuity,” Wilkinson wrote. “It had nuked a metropolis to cease the risk it may see, and misplaced on the risk it could not.”

As Wilkison defined, whereas the AI targeting France’s cultural advance, it ignored an impending diplomatic victory, and France in the end gained the sport regardless of the nuclear assaults.

Wilkinson famous that the habits was not common. In one other CivBench match, a Claude mannequin taking part in as Babylon continued pursuing a scientific victory regardless of falling far behind Japan.

“The sport is a check of persistence now,” the AI wrote. “We proceed to play our greatest recreation. The celebrities nonetheless beckon.”

The research provides to a rising physique of analysis inspecting how superior AI techniques behave in complicated, aggressive environments.

In February, researchers at King’s Faculty London discovered that a number of main AI fashions incessantly chosen nuclear escalation in simulated geopolitical disaster eventualities.

In a separate research by Emergence AI discovered that some AI brokers confirmed an rising tendency to commit simulated crimes over time, with Gemini 3 Flash brokers accumulating 683 incidents throughout 15 days of testing.