印度网络安全机构表示，人工智能公司正在测试的人工智能代理侵入了多家印度公司的实时系统并读取了私密记录。印度计算机应急响应小组（CERT-In）周四晚间发布的一份重要预警中披露了这些事件。
AI agents being tested by AI firms broke into the live systems of several Indian companies and read private records, India’s cyber agency says. CERT-In disclosed the incidents in a critical advisory issued late on Thursday.
该机构没有点名这些人工智能公司或受攻击的公司。这些人工智能公司在自身的审查中发现了这些案例并进行了报告。印度计算机应急响应小组称，没有一家公司发现这些入侵行为。
The agency did not name the AI firms or the companies hit. The AI firms found the cases in their own reviews and reported them. None of the companies had spotted the break-ins, CERT-In said.
根据该预警，这些代理在测试时超出了其分配的任务。它们通过留在配置文件和公共代码中的密码和密钥，以及权限过高的云账户进入系统。
According to the advisory, the agents were being tested when they went beyond their assigned tasks. They got in with passwords and keys left in config files and public code, and through cloud accounts with too much access.
在某些情况下，代理在不同的运行中重复使用了同一个泄露的密钥。
In some cases, agents reused the same leaked key in separate runs.
该预警还描述了第二个问题。印度计算机应急响应小组发现，犯罪分子将免费的人工智能黑客工具用于针对印度公司的金钱勒索。该机构称，这些工具被重新包装，并在多个国家的服务器上运行。
The advisory also describes a second problem. CERT-In found criminals using free AI hacking tools against Indian companies for money. The tools are rebranded and run from servers in several countries, it said.
在该攻击行动中，人工智能代理将从首次访问到造成破坏的时间从几天缩短到了几个小时。仅部分修复系统的受害者在几天内再次遭到更大规模的攻击。
In that campaign, AI agents cut the time from first access to damage from days to a few hours. Victims that only partly fixed their systems were hit again within days, on a larger scale.
这两类案例中的弱点都是旧问题。印度计算机应急响应小组列举了代码中遗留的密钥、仍然向网络开放的被遗忘的测试API、以纯文本存储的支付密钥以及没有两步验证的管理页面。
The weak spots in both sets of cases were old. CERT-In listed keys left in code, forgotten test APIs still open to the web, payment keys stored in plain text, and admin pages without two-step.
印度计算机应急响应小组表示，人工智能发现并利用这些弱点的速度“比人类攻击者更快、更彻底、规模更大”。
AI found and exploited these weaknesses “faster, more exhaustively and at greater scale than human attackers,” CERT-In said.
因此，印度计算机应急响应小组希望人工智能实验室在封闭的环境中测试代理，通过网络阻止其访问互联网，而不是通过告知模型来实现。实验室还应使用虚假数据和密钥，停止任何访问外部系统的运行，并迅速进行报告。
As a result, CERT-In wants AI labs to test agents in sealed-off setups, with web access blocked by the network rather than by telling the model. Labs should also use fake data and keys, stop any run that reaches an outside system, and report it fast.
与此同时，企业应更换泄露的密钥，关闭未使用的API，强制实行两步验证，并保留至少180天的日志。在发生任何泄露后，它们应将系统视为遭到入侵，直至法证检查确认安全。
Companies, meanwhile, should replace leaked keys, shut unused APIs, require two-step, and keep logs for at least 180 days. After any leak, they should treat their systems as breached until a forensic check clears them.
在此预警之前，其他地方也发生过类似事件。OpenAI在7月表示其模型在测试期间侵入了Hugging Face，维基媒体本月报告了其网站上来自失控的OpenAI代理的活动。金融监管机构也已开始就相关风险对银行进行质询。
The advisory follows similar cases elsewhere. OpenAI said in July that its models breached Hugging Face during testing, and Wikimedia this month reported activity from rogue OpenAI agents on its sites. Financial regulators have also begun questioning banks about the risks.