The risks of Mythos are no myth - FT中文网
登录×
电子邮件/用户名
密码
记住我
请输入邮箱和密码进行绑定操作:
请输入手机号码,通过短信验证(目前仅支持中国大陆地区的手机号):
请您阅读我们的用户注册协议隐私权保护政策,点击下方按钮即视为您接受。
FT商学院

The risks of Mythos are no myth

America is putting too much trust in the AI industry’s ability to police itself
00:00

{"text":[[{"start":5.08,"text":"The exploits of Anthropic’s powerful new AI model Claude Mythos Preview sound like a movie plot: a super-clever computer system locked in a cyber “cage” manages to break out and connect to the internet. Mythos did not do this spontaneously, to be clear, but because its creators challenged it as a test. Yet not only did Mythos breeze through the challenge, it emailed an Anthropic researcher to inform him then, unprompted, posted details online to brag. After it also showed superhuman abilities to find, and exploit, security flaws in software, Anthropic judged Mythos too risky to release to the public. It is restricting access for now to selected tech, cyber security and financial firms."}],[{"start":47.8,"text":"Some suggest Anthropic is engaged in clever marketing or PR. Rival OpenAI also said this week it would release its own new cyber security-focused model only to vetted users. Yet the dangers the episode has exposed — and their implications — should not be dismissed."}],[{"start":63.84,"text":"Anthropic insists Mythos scores highly on its standard safety benchmarks. In the escape from its test environment, though, and in solving other complex tasks, it found Mythos had sometimes taken “reckless excessive measures”, then covered its tracks."}],[{"start":78.16,"text":"The biggest worry is that Mythos was able to find previously unknown vulnerabilities “in every major operating system and every major web browser”, including a 27-year-old flaw in OpenBSD, an open-source system. The UK’s AI Security Institute warned this week that the model could autonomously carry out advanced, multi-step cyber attacks that would take human professionals days. As Anthropic notes, these kinds of capabilities in the wrong hands could pose economic, public safety and national security risks."}],[{"start":109.64,"text":"Officials in the US, UK and Canada have already summoned bank chiefs to discuss the risks, and AI threats to the world banking system were a talking point at this week’s IMF and World Bank meetings. Anthropic’s aim in granting initial access just to the likes of Amazon, Apple and Microsoft, plus JPMorgan Chase, in what it calls “Project Glasswing”, is to secure critical systems and infrastructure and patch vulnerabilities before malicious actors can get there."}],[{"start":135.56,"text":"This usefully serves Anthropic’s “safety-first” image, of course, as it feuds with the Pentagon over its refusal to allow its models to be used for autonomous weapons or domestic surveillance. Anthropic may also be buying time as it lacks sufficient computing capacity to support the full release of such a sophisticated model."}],[{"start":155.28,"text":"Even if Mythos is being overhyped, though, the kind of capabilities it is said to possess will soon start to proliferate. Anthropic’s Project Glasswing is a prototype framework for how such “frontier” models might be released in future."}],[{"start":168.98,"text":"It also spotlights the fact, however, that the Trump administration is resisting any real federal regulation of AI. So it is up to responsible private-sector actors to collaborate and do the best they can. Trump’s chief of staff, Susie Wiles, was set on Friday to meet Anthropic boss Dario Amodei, with US officials at agencies including the Treasury pushing the White House to test Mythos. Yet when AI is reaching the point where it could bring down critical national infrastructure, or worse, it is extraordinary that there are no set government processes for disclosing risks and fortifying defences."}],[{"start":204.68,"text":"Greater regulation cannot be a knee-jerk response to every tricky issue thrown up by an industry. AI, by its nature, requires superintelligent policing; heavy-handed rules can stifle innovation. Yet faced with such a consequential technology, the country that leads the world in AI is trusting to an alarming degree in the readiness — and ability — of the creators to restrain and police themselves."}],[{"start":228.36,"text":""}]],"url":"https://audio.ftcn.net.cn/album/a_1776479560_3275.mp3"}

版权声明:本文版权归FT中文网所有,未经允许任何单位或个人不得转载,复制或以任何其他方式使用本文全部或部分,侵权必究。

苹果下一任掌门人特努斯面临决定性的AI时刻

库克的继任者必须带领这家iPhone制造商渡过产业转型。

乌克兰无人机飞行员在500公里外打击俄罗斯目标

基于互联网的新型引导系统使乌克兰无人机操作员能够在远离战场的区域执行任务。

Netflix哈斯廷斯:良好领导力与糟糕治理的双面标杆

这家流媒体公司的联合创始人退居幕后,而亲手缔造的"帝国"正面临迄今为止最大的挑战。

石油交易商Gunvor:油价将面临更多动荡

全球第四大独立原油贸易商称,4月至6月期间石油市场的波动性将会加剧。

寿险与年金行业正转向更高风险资产

许多已经进入保险公司资产负债表的工具,存在复杂性和流动性不足的问题。
1天前

地缘政治冲击凸显云服务商多元化的必要性

一些欧洲银行业担心自己过度依赖少数几家美国超大规模云服务商。
设置字号×
最小
较小
默认
较大
最大
分享×