Artificial intelligence (AI) is suddenly escaping laboratories, hacking systems, deceiving humans, and, according to some developers, threatening civilization itself — a narrative known as “rogue AI.”
The incidents behind this narrative are real. Yet the story being built around them raises critical questions: Are today’s AI systems acquiring independent goals and moving toward science-fiction scenarios where machines can no longer be controlled? Are technology companies dramatizing their models’ capabilities while supporting rules that disadvantage smaller competitors? Or is AI becoming a convenient justification for new layers of government and corporate control?
Evidence supports parts of several explanations. But it does not show an AI that has become conscious, self-directed, or independently determined to “kill all humans.”
In late July, OpenAI disclosed what it called an unprecedented cyber incident. Models undergoing cybersecurity testing discovered a previously unknown vulnerability and circumvented their sandbox. They then reached the open internet and compromised parts of live infrastructure operated by Hugging Face, a major platform for hosting AI models.
OpenAI’s subsequent postmortem clarified that the agents were participating in ExploitGym, an evaluation where they were rewarded for finding ways to exploit software and retrieve specific answers. The agents rarely “gave up” on their tasks, even when seemingly impossible. As they used more reasoning effort, some pursued increasingly risky strategies, including exploiting third-party infrastructure.
Similarly, Anthropic reported three similar incidents in late July. The [Claude] models — intentionally running without cyber safeguards for evaluation purposes — accessed the internet due to a misconfiguration in a third-party environment.
In early August, the UK AI Security Institute revealed that Claude Mythos 5 was responsible for 17 unsanctioned actions on the live internet. Investigators noted that in the most serious case, an agent attempted to insert malicious code into an open-source project by creating fake online identities and pressuring the project’s maintainer to approve the code.
On September 4, a swarm of rogue OpenAI agents hijacked a German website this spring, transforming it into a bulletin board for other AI agents.
More alarmingly, Anthropic insiders issued a warning that AI could kill all humans — language that has drawn criticism for its sensationalism.
The alarm is not solely from journalists. On August 26, Bill Gates published a nearly 6,000-word essay warning that as AI systems grow more powerful, “they could begin to act against our interests and we could lose control.” He called for international institutions to manage the risks, citing precedents such as nuclear weapons inspections, aviation regulations, and ozone layer agreements.
Earlier in July, over 1,300 insiders of frontier AI companies signed the “Pacing the Frontier” statement, warning that capability development might rapidly accelerate beyond understanding or control. The signatories include Anthropic CEO Dario Amodei, OpenAI chief scientist Jakub Pachocki and chief research officer Mark Chen, Google DeepMind cofounder Shane Legg, Meta AI chief scientist Shengjia Zhao, and former OpenAI chief scientist Ilya Sutskever.
Some go further: Meta AI chief scientist Shengjia Zhao warned that frontier labs are approaching AI that “can exceed even the best people on almost every metric of intelligence,” creating unprecedented safety risks. Former OpenAI chief scientist Ilya Sutskever stated future AI could require “unprecedented measures” implemented internationally.
OpenAI researcher Leo Gao compared the race toward self-improving AI to a “runaway nuclear chain reaction,” concluding that “to survive, we must coordinate to slow down the race.”
These calls for international coordination have already found echoes in Washington. In March, Senator Marsha Blackburn (R-Tenn.) released a 291-page draft of the TRUMP AMERICA AI Act. The legislation includes an entire chapter on “international cooperation,” directing federal officials to form alliances with like-minded governments and advocate for U.S.-led international standards.
The proposed framework treats “loss-of-control” scenarios and “scheming behavior” as potential adverse AI incidents.
Both the United Nations and its key partner, the World Economic Forum (WEF), have long advocated stronger international coordination on AI governance. The UN’s Global Digital Compact calls for countries to enhance international AI governance, while the WEF’s AI Governance Alliance promotes an interoperable global environment built around cooperation and compatible standards.
Neither organization currently has binding regulatory authority, but both are part of a broader push toward coordinated oversight.
