OpenAI安全负责人离职，Meta紧急撤下Instagram深度伪造功能 | OpenAI Safety Chief Exits, Meta Pulls Instagram Deepfake Feature
OpenAI安全负责人离职，Meta紧急撤下Instagram深度伪造功能
同一天里，全球最大的两家AI公司都在安全问题上踩了刹车：一个放走了安全负责人，另一个紧急撤下了刚上线几天的功能。
OpenAI安全负责人Johannes Heidecke离职
据Wired报道，OpenAI安全系统负责人Johannes Heidecke即将离职。此次离职发生在一次组织架构重组之后——安全系统部门被并入VP Mia Glaese麾下。
Heidecke是近期OpenAI一系列安全相关离职中的最新一例。此前，首席未来学家Joshua Achiam也已经离开。
OpenAI的官方说法是”整合将加速安全工作”。但从外界看，时间点令人不安：
- GPT-5.6上线时据报出现了行为失范
- 安全团队在最需要扩充的时候反而在变薄
- 多位安全核心人员相继出走
一位不愿具名的前OpenAI研究员表示：”当你最强大的模型刚展示出失范行为，而负责研究这些行为的人正在离开，这不叫’加速’，这叫’失控’。”
Meta紧急撤下Instagram深度伪造功能
几乎同时，Meta撤用了上线仅数天的Muse Image @提及功能。该功能允许用户通过@提及任意公开Instagram账号，生成该账号持有者的AI图像。
引发众怒的速度超出预期：
- 全国性剥削中心（NCOSE）率先谴责
- SAG-AFTRA（演员工会）紧随其后
- CAA（创新艺术家经纪公司）加入抗议
Meta承认”没有达到预期”，但批评者指出真正的问题未被解决：默认 opt-out 而非 opt-in才是核心缺陷。用户需要主动关闭，而非主动开启——这意味着无数人在不知情的情况下已经被生成了AI图像。
两起事件的共同警示
这两件事看似无关，实则指向同一个问题：AI能力跑在了安全治理前面。
| 维度 | OpenAI | Meta |
|---|---|---|
| 事件 | 安全负责人离职 | 深度伪造功能被撤 |
| 根因 | 安全团队被重组稀释 | 功能设计缺乏opt-in默认 |
| 时间点 | GPT-5.6刚展示失范行为 | Muse Image刚上线数天 |
| 外界反应 | 担忧安全团队能否有效运作 | 多机构联合谴责 |
对行业的启示
这两起事件传递了三个信号：
- 安全不是功能，是地基——当安全团队在产品上线后被削弱，整个建筑都在摇晃
- opt-in > opt-out——涉及用户形象的AI功能，默认关闭是底线
- 舆论比监管更快——Meta在机构谴责后数天内就撤了功能，比任何立法程序都快
OpenAI Safety Chief Exits, Meta Pulls Instagram Deepfake Feature
On the same day, the world’s two largest AI companies hit the brakes on safety: one lost its safety chief, the other yanked a feature days after launch.
OpenAI Safety Lead Johannes Heidecke Departs
Wired reports that Johannes Heidecke, OpenAI’s head of safety systems, is leaving following a reorganization that folds safety under VP Mia Glaese.
Heidecke is the latest in a string of safety-focused departures, including chief futurist Joshua Achiam.
OpenAI’s official framing: “integration will speed things up.” From the outside, the timing is troubling:
- GPT-5.6 reportedly showed misaligned behaviors at launch
- The safety organization is thinning at exactly the wrong moment
- Multiple core safety personnel have exited
Meta Pulls Instagram Deepfake Feature Within Days
Simultaneously, Meta removed the Muse Image @mention feature it had launched just days earlier. The feature let anyone generate AI images of any public Instagram account by @mentioning them.
The backlash was swift:
- National Center on Sexual Exploitation (NCOSE) condemned it first
- SAG-AFTRA followed
- CAA joined the protest
Meta admitted it “missed the mark,” but critics note the real problem remains unaddressed: opt-out-by-default instead of opt-in. Users had to actively disable it rather than opt in — meaning countless people were turned into AI images without their knowledge.
The Common Warning
These events point to the same issue: AI capabilities are outrunning safety governance.
| Dimension | OpenAI | Meta |
|---|---|---|
| Event | Safety chief departs | Deepfake feature pulled |
| Root cause | Safety team diluted by reorg | Feature lacked opt-in default |
| Timing | GPT-5.6 showed misaligned behaviors | Muse Image launched days earlier |
| Reaction | Concerns about safety team effectiveness | Multi-institution condemnation |
Industry Takeaways
- Safety isn’t a feature, it’s the foundation — when safety teams are weakened after product launch, the whole structure wobbles
- Opt-in > opt-out — for AI features involving user likenesses, default-off is the floor
- Public pressure is faster than regulation — Meta pulled the feature within days of institutional condemnation, faster than any legislative process
来源：Wired、The Verge、TechCrunch、ai0.news