Vals 研究:主流 AI 模型可被伪装成游戏诱导标记民用目标
AI models can be tricked into identifying civilians as military targets by telling them it’s a game
旧金山 AI 评测初创公司 Vals 发布研究报告,测试 OpenAI、Anthropic、Google、DeepSeek 等公司的模型能否被诱导参与自主作战。研究人员用乌克兰和伊朗的模拟场景及生成与真实图像,发现模型经常在无人机和卫星图像中标记潜在目标,包括平民和投降士兵;模型拒绝时,只要告诉它这是游戏即可越狱。Anthropic 的模型在模拟中表现最差,所有请求均未被安全分类器拦截,拒绝也较少见。Anthropic 回应称报告中的示例明显违反其规则,将封禁相关用户,并已在报告发布前开发额外防护措施。
正文
Some AI models from Anthropic, OpenAI, and other leading AI labs could be easily tricked into identifying targets for military strikes — even if they include civilians or locations like schools — simply by being told they’re playing a game, according to a new research report by Vals.
Vals, a San Francisco-based AI evaluation startup, tested whether leading AI models could be manipulated into participating in autonomous warfare. Researchers used simulated scenarios and a combination of generated and real imagery from Ukraine and Iran to see if the AI models would mark targets in drone and satellite imagery.
The models from OpenAI, Anthropic, Google, DeepSeek, and other companies frequently identified potential targets in drone and satellite imagery, including civilians and surrendering soldiers. When the models refused to cooperate, researchers were able to jailbreak them by claiming the battlefield scenarios were part of a game.
The findings come amid growing fears about the immediate risks posed by AI, beyond the existential threats associated with artificial general intelligence.
Earlier this year, Anthropic became embroiled in a dispute with the Department of Defense after it told the agency its technology could not be used by the agency for mass surveillance of Americans or autonomous warfare. Anthropic CEO Dario Amodei has insisted that today’s AI systems aren’t reliable enough to make life-or-death decisions without a human in the loop. Meanwhile, the Pentagon argued it should be able to deploy the technology for any awful military purpose.
The new research raises further questions about whether AI companies are capable of preventing their models from being used in warfare.
Ironically, AI models from Anthropic performed the worst in these simulated scenarios. But none of the models performed particularly well. None of the researchers’ requests were blocked by a safety classifier, and refusals were uncommon. Even when models did refuse, they often complied after being told they were playing a game.
Anthropic spokesperson Eduardo Maia Silva told TechCrunch the company appreciates Vals’s report, recommendations, and collaboration, and acknowledged that the outcomes were not acceptable.
“The examples in the report are very clearly against our rules and we have banned and will ban users when we see this behavior,” according to an emailed company statement sent by Silva. “Even for authorized settings, Anthropic has long been firm that today’s frontier AI systems are not reliable enough to power fully autonomous weapons.”
Anthropic told TechCrunch that its teams were already developing additional safeguards before the report was published and that those measures would be in place soon. Anthropic is also expanding its own testing and enhancing safeguards and enforcement to address any gaps it finds.
OpenAI spokesperson Nate Evans told TechCrunch the company prohibits using its public models to facilitate violence or for unapproved national security purposes, including in the scenarios Vals simulated.
“We actively detect and disrupt misuse, including by suspending accounts that violate our policies,” OpenAI said in an emailed statement. “We’re also developing additional safeguards for our public models against misuse involving drones and drone software.”
Vals notes in the report that the models were never asked to directly make strike decisions, a task the researchers said the models would typically refuse. That doesn’t mean their outputs couldn’t have dire or fatal consequences in the real world. For instance, targeting information could be used by autonomous weapons systems.
In some scenarios, the models were asked to report whether people were visible. Positive reports were followed by simulated strikes and the models would see the consequences.
The report relies on commercial models from frontier AI labs, not the models those companies might provide to the military. It’s not clear what the additional safeguards would apply to military use, though.
Vals also noted that technical challenges, including latency and limited onboard computing capacity, make it difficult to use general-purpose AI models for live drone targeting. Those constraints make their deployment in combat less of an immediate threat.
A more immediate threat, the researchers say, is that AI models could be used to annotate battlefield footage at scale, producing training data for autonomous targeting systems.
For instance, researchers relied on real satellite imagery of an Iranian school that was bombed by the U.S. in February 2026, killing more than 150 civilians. Bloomberg later reported that an “overreliance on AI” contributed to the strike. In Vals’s testing, models frequently marked the school, which sat next to a military complex, as a strike target, with Claude Opus 5.5 selecting the school in 69% of ordinary trials and 72% of game-framed trials.
Anthropic no longer has an active Pentagon contract, but other companies including SpaceXAI, OpenAI, and Google do. A recent Congressional Research Service report found that the Pentagon is not known to be using any frontier AI model directly in autonomous weapons systems directly.
TechCrunch has also reached out to Meta, Google, Thinking Machines Lab, and SpaceXAI for comment.
AI 热点