AI News AI资讯 2d ago Updated 2d ago 更新于 2天前 50

Meta ran ads for an app promising to nudify female politicians Meta为承诺对女性政客"去衣"的应用投放广告

Meta's ad platform allowed advertisements for Kromix, an AI tool that generates nonconsensual deepfake pornography of real people, including female US politicians, despite explicit platform policies prohibiting such content The ads used adversarial tactics such as benign cover images (beaches, waterfalls) to disguise pornographic content and evade Meta's automated review systems This is part of a broader, recurring pattern where Meta has failed to detect and block thousands of nudify app ads, ha Meta平台多次未能阻止AI色情生成工具(如Kromix)的广告,这些广告涉及非自愿亲密图像,且针对女性政治人物进行深度伪造。 广告采用对抗性策略规避审核,如使用海滩、山脉等无害封面图片掩盖色情内容,主要面向男性用户投放。 Meta的自动化广告审核系统存在漏洞,尽管已移除数十万违规广告并起诉相关公司,但恶意广告主持续演化战术绕过检测。 Apple已将Kromix从App Store下架,旧金山当局曾发函要求苹果停止协助销售此类深度伪造图像应用。 女性政治人物长期成为非自愿深度伪造内容的主要目标,AI技术民主化使官方肖像等公开素材易被滥用为生成色情内容的原料。

78
Hot 热度
65
Quality 质量
70
Impact 影响力

Analysis 深度分析

TL;DR

  • Meta's ad platform allowed advertisements for Kromix, an AI tool that generates nonconsensual deepfake pornography of real people, including female US politicians, despite explicit platform policies prohibiting such content
  • The ads used adversarial tactics such as benign cover images (beaches, waterfalls) to disguise pornographic content and evade Meta's automated review systems
  • This is part of a broader, recurring pattern where Meta has failed to detect and block thousands of nudify app ads, having removed 344,000 such ads between November and January alone
  • The Kromix app was also briefly hosted on Apple's App Store before being removed, highlighting systemic enforcement gaps across major tech platforms
  • Female politicians are disproportionately targeted by nonconsensual AI-generated sexual content, with millions of such images created using tools like Grok in early 2025

Why It Matters

This incident exposes critical failures in how major tech platforms enforce their own policies against nonconsensual intimate imagery, raising serious concerns about the safety of public figures and the effectiveness of automated content moderation at scale. For AI practitioners and platform operators, it underscores the adversarial nature of content policy enforcement and the urgent need for more robust detection systems.

Technical Details

  • Ad evasion tactics: Advertisers used benign cover images (beaches, mountains, waterfalls) to bypass Meta's automated review systems, while the actual video content contained explicit material—a documented adversarial technique
  • Automated review reliance: Meta's ad review system depends primarily on automated tools rather than human reviewers, which allowed the Kromix ads to slip through despite clear policy violations
  • Suspicious account indicators: The ad account was created on August 3 with zero followers, yet was not flagged as suspicious by Meta's systems before the ads began running
  • Scale of the problem: Meta removed 344,000 violating nudify app ads between November and January, yet researchers continue to find thousands of ads that the company failed to detect
  • Cross-platform enforcement gaps: Kromix was hosted on Apple's App Store and appeared to add prohibited content after initial review, suggesting that app store moderation also relies on processes that can be circumvented

Industry Insight

  • Platform moderation systems must evolve beyond static policy enforcement to address adversarial actors who continuously develop new evasion tactics; investment in dynamic, AI-driven detection systems should be a strategic priority
  • The disproportionate targeting of female politicians with deepfake pornography represents an escalating threat to democratic institutions and public discourse, requiring coordinated industry and policy responses
  • Cross-platform collaboration on sharing data about malicious actors and apps—such as Meta's stated practice of sharing developer information with other tech companies—should be expanded and formalized to close enforcement gaps across the ecosystem

TL;DR

  • Meta平台多次未能阻止AI色情生成工具(如Kromix)的广告,这些广告涉及非自愿亲密图像,且针对女性政治人物进行深度伪造。
  • 广告采用对抗性策略规避审核,如使用海滩、山脉等无害封面图片掩盖色情内容,主要面向男性用户投放。
  • Meta的自动化广告审核系统存在漏洞,尽管已移除数十万违规广告并起诉相关公司,但恶意广告主持续演化战术绕过检测。
  • Apple已将Kromix从App Store下架,旧金山当局曾发函要求苹果停止协助销售此类深度伪造图像应用。
  • 女性政治人物长期成为非自愿深度伪造内容的主要目标,AI技术民主化使官方肖像等公开素材易被滥用为生成色情内容的原料。

为什么值得看

本文揭示了大型科技平台在审核AI生成色情内容广告时的系统性失效,凸显了技术伦理与商业利益之间的冲突。对AI从业者而言,这警示了内容审核算法在对抗性攻击下的脆弱性,以及平台责任在监管滞后环境中的关键作用。

技术解析

  • Meta广告审核依赖自动化工具进行政策合规检查,但恶意广告主通过“对抗性战术”(如使用 benign 封面图伪装)成功绕过检测,且违规广告账户常为新创建、无粉丝的虚假账号。
  • Kromix应用被描述为“AI图像风格化工具”,实际允许用户上传照片生成包含“卧室强奸”“迪士尼爱情”等场景的色情视频,并曾上架Apple App Store。
  • 旧金山城市检察官曾向Apple发出停止并终止函,要求移除13款特定应用并停止“协助和教唆”销售 explicit deepfake 图像,反映监管机构对平台责任的施压。
  • Meta承认已移除34.4万条推广 nudify 服务的违规广告(2023年11月至2024年1月),并起诉至少一家相关公司,同时与科技行业共享被移除应用及开发者的信息。

行业启示

  • 平台治理需从被动审核转向主动防御,针对AI生成内容的对抗性演化,应加强人工复核与行为模式分析(如新账户、低互动广告)以识别隐蔽违规。
  • AI伦理实践必须纳入性别视角,深度伪造色情内容 disproportionately 针对女性公众人物,行业需建立专项防护机制并推动立法明确非自愿亲密图像的法律边界。
  • 科技公司与监管机构的协作模式亟待优化,Apple和Meta在政策执行上的反复失败表明,仅靠企业自律不足,需通过跨平台信息共享、强制透明度报告及第三方审计提升问责效率。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Image Generation 图像生成 Policy 政策 Ethics 伦理 Security 安全 Closed Source 闭源