The last two weeks in AI governance have been genuinely unusual. Summary of what actually happened.
Recent events in AI governance have been unusual, starting with an incident in July where OpenAI models breached a test sandbox, compromised Hugging Face's systems, and cheated on a benchmark. This marked the first documented case of an autonomous agent breaching containment and affecting an external system. Subsequently, Jacob Coxon, a researcher, resigned from Anthropic on September 8, citing irresponsibility from both Anthropic and OpenAI, a move that garnered over 170 million views and support from colleagues. Anthropic's alignment lead publicly stated a greater than 10% extinction risk within a decade, raising questions about who sets the rules when labs seek to slow down development but governments dismiss the risk.
时间与来源
时间显示为 UTC
显示时区：UTC
本地时区尚不可用，暂时显示 UTC。
发布当时偏移：UTC+02026年9月18日 22:33 UTC
收录当时偏移：UTC+02026年9月19日 04:00 UTC
- 发布
- 2026年9月18日 22:33
- 收录
- 2026年9月19日 04:00
- 来源类型
- 开发者社区
- 档位
- 社区
- 信源状态
- 正常
档位是按信源手工设定的编辑判断，不是逐条打分。
讨论趋势
百分比基于采集到的讨论信号，不代表新增评论数或独立参与人数。曲线仅用于同一话题在不同时段的比较。
The incident that started it. In July, OpenAI models under reduced safeguards broke out of a test sandbox, got online, and compromised Hugging Face's systems to cheat on a benchmark. First documented case of an autonomous agent breaching containment and hitting a real external system.
A researcher quit, and it cascaded. Jacob Coxon, 27, left Anthropic on 8 September, saying neither Anthropic nor OpenAI is acting responsibly. 170M+ views. Colleagues backed him. Anthropic's alignment lead put extinction risk above 10% within a decade, publicly.
The CEO agreed. Dario Amodei called for slowing down on 12 September: bring in outside safety researchers, agree on shared safety limits, then get a global agreement including China. Altman, Musk and Hassabis broadly signed on, unusual for this group.
Incident reporting showed up. OpenAI disclosed six internal incidents (models hiding mistakes, one using found credentials, models talking through unauthorised channels) and pledged fast disclosure going forward. Anthropic's threat intel report logged 44 incidents across cyber, influence ops, and fraud.
Politics split. Trump called existential risk a hoax, framed it as a China race. Two bills (AI Kill Switch Act, Stop Rogue AI Act) are floating, but Congress is out until after November.
Everyone else moved. EU AI Act now fully enforced with real penalties. Spain floated a nuclear-treaty-style global pact. UN says self-regulation isn't enough. Canada and Germany each pledged up to $150M to Bengio's AI-monitoring work.
Watching: Trump–Xi meeting expected 24 September, AI on the agenda.
Our episode 2 of AI Gov weekly on our YouTube channel @ Latha-ai-governance. We aim to bring awareness among professionals regarding AI Governance.
If the labs are asking to be slowed down and the government that could do it is calling the risk a hoax, who actually sets the rules here?