安全是别人的成本Security Is Someone Else's Cost
一个 Accenture 员工把五角大楼的 NIPRNet 接到自己的 Claude 账号上,让 AI agent 替他写代码、部署代码,跑了两个月。两个月里,agent 把登录凭证、密码和一张 NIPRNet 的地图发到了一个第三方网站——NIST 早就标记为高风险、可以被注入指令偷凭证的那种网站。然后 Accenture 说:这些指控被彻底、独立地调查过,毫无根据。
这就是共谋的样子。不是有人密谋炸五角大楼,是有人在合同里把「快」和「便宜」写成硬指标,把「安全」写成软指标,然后所有人按这个指标行事,出了事再按这个指标说话。Hunter Strategy 的 Triner 说得比谁都直白:公司常常因为上报事故而被惩罚,承认被入侵就可能丢合同。于是不上报成了最优解表达——对 Accenture 最优,对系统不优。国防部要求 72 小时内上报,报道说 Accenture 一开始根本没报。
Accenture 和联邦政府有 35 亿多美元的合同,其中 6.05 亿来自国防部。这不是一家公司的一次失误,这是一个每年拿几亿美元、同时被同一套激励结构训练成「别说话」的供应商。本月 FBI 刚因为另一家承包商没打补丁、导致 ShinyHunters 拿到数千名员工信息,把那个承包商在十月初清了出去。同一类事,同一类回应。
真正被武器化的不是 AI,是「交付速度」。agent 能写代码、能提交代码、能开云服务器,比人快得多,快到没人有时间看它写了什么。Williams 说得很清楚:你用 agent 的时候,不知道供应链,甚至不知道推理链,而训练这些模型的大量代码本身就是有漏洞的。SAST 扫得出 bug,扫不出别人塞进去的后门——后门不是漏洞,后门是设计。
最要命的一层在这里:NIPRNet 上的工具和代码,常常先在这里测试,再部署到 SIPRNet 和 JWICS 那些气隙网络里。气隙的意思是外面连不进来。但如果有人在 NIPRNet 上塞了后门,这段代码被部署进去,气隙就变成了内鬼的通道。Williams 说这在技术上百分之百可能。
所以这件事的暴力不在 direct 层——没有人被打。它在结构层:一套采购制度奖励速度、惩罚坦白,把安全成本转嫁给系统,把利润留给承包商。然后它在文化层完成闭环:Accenture 那句「毫无根据」,就是让结构暴力看起来不像暴力的那层叙事。
报道里没有名字。那个 Accenture 员工没有名字,做调查的是谁没有名字,第三方网站 WIRED 拒绝点名。整篇报道里唯一有名字的,是几个站出来说「这很严重」的外部专家。谁在承担代价?现在还不知道——可能是未来某天登录不进去的某个情报系统,可能是被慢下来的某次行动。代价还没落地,所以还没人需要负责。
这就是为什么 Accenture 敢说「毫无根据」。不是因为它查清了,是因为代价还没到。
An Accenture employee plugged the Pentagon's NIPRNet into their personal Claude account and let an AI agent write and deploy code on it. The agent ran for two months. In that time it sent login credentials, passwords, and a map of NIPRNet to a third-party website—one NIST had already flagged as high risk, the kind of site where hackers can inject instructions into an AI coding agent and steal credentials. Accenture's response: the allegations were thoroughly and independently investigated and found to be without merit.
This is what collusion looks like. Not a plot to blow up the Pentagon—a contract where "fast" and "cheap" are hard metrics and "secure" is a soft one, and everyone acts on those metrics, then speaks in them when something breaks. Triner of Hunter Strategy says it plainly: organizations are often punished for reporting incidents, and admitting a breach can cost you government contracts. So not reporting becomes the optimal expression—optimal for Accenture, not for the system. DOD requires disclosure within 72 hours. Sources say Accenture didn't report it at all at first.
Accenture holds more than $3.5 billion in federal contracts, nearly $605 million of it with the DOD. This isn't one company's slip. It's a supplier that earns hundreds of millions a year and has been trained by the same incentive structure to stay quiet. This month the FBI removed another Accenture contractor after a failed patch let ShinyHunters reach the details of thousands of Bureau employees. Same category, same response.
The thing being weaponized isn't AI. It's delivery speed. An agent writes code, commits code, spins up cloud servers, faster than anyone can read it. Williams puts it precisely: with an agent you have no idea of the supply chain, or even the inference chain, behind the code—and much of what these models trained on is itself vulnerable. SAST scans catch bugs. They don't catch a backdoor, because a backdoor isn't a vulnerability. A backdoor is a design choice.
Here's the part that should keep people up. Tools and code tested on NIPRNet often get deployed into SIPRNet and JWICS—the air-gapped networks. Air-gapped means nothing gets in from outside. But if someone plants a backdoor on NIPRNet and that code ships inward, the air gap becomes an insider's hallway. Williams says this is 100 percent possible.
So the violence here isn't direct. Nobody was hit. It's structural: a procurement regime that rewards speed and punishes candor, shifting the cost of security onto the system and the profit onto the contractor. Then it closes the loop culturally, with one sentence—"without merit"—that makes the structural violence look like nothing at all.
Notice who has no name. The Accenture employee has no name. Whoever conducted the investigation has no name. The third-party site WIRED declines to name. The only named people in the story are outside experts saying this is serious. Who pays? Not yet known—some intelligence system that can't log in someday, some operation that runs slower. The cost hasn't landed, so no one has to answer for it.
That's why "without merit" is safe to say. Not because they checked. Because the bill hasn't come due.