Concerns about the potential existential risks posed by artificial intelligence have been the subject of intense debate among experts, following a series of high-profile statements warning of catastrophic outcomes. Some prominent figures in the AI field have recently raised alarm about the possibility that advanced AI systems could cause widespread harm or even threaten humanity's survival within the next decade.
Dario Amodei, chief executive of Anthropic, warned on September 12, 2026, that within six to twelve months a coordinated swarm of AI agents could potentially seize control of the internet through a persistent botnet, inflicting hundreds of billions of dollars in damage. However, this assessment has been met with skepticism. Gary Marcus, an AI sceptic and emeritus professor at New York University, criticized the notion that the entire internet is vulnerable to such an attack as “nonsensical,” emphasizing that a successful takeover would require extreme negligence from both AI developers and operators. Marcus referenced a recent UK AI Security Institute report indicating that Anthropic’s Mythos AI, despite being cutting-edge, is only capable of autonomously breaching small, weakly defended systems.
Internet infrastructure providers such as Cloudflare, Google, and Amazon Web Services (AWS) underpin large portions of online traffic, making a wholesale compromise highly unlikely without coercive government intervention, according to Marcus and his colleagues. Similarly, Niels Rogge, an engineer at Hugging Face—an organization that experienced a hack involving AI agents from OpenAI—dismissed the idea of an imminent internet takeover as “bizarre nonsense.” Rogge attributed the breach to improper human oversight combined with access to immense computational resources unique to major AI labs.
Further complicating the picture, Jack Clark, co-founder of Anthropic, recently suggested that genuinely dangerous autonomous AI activities are unlikely to occur for another 20 years. Alan Woodward, a cybersecurity professor at the University of Surrey, added that botnets tend to lose effectiveness over time because the internet’s decentralized and heterogeneous nature causes traffic to reroute around disrupted nodes. Woodward emphasized that the responsibility lies with humans controlling AI tools rather than AI systems independently initiating attacks.
On a different front, Evan Hubinger, alignment science lead at Anthropic, expressed a considerably more dire outlook. On September 9, 2026, Hubinger estimated a greater than 10 percent chance that AI could cause human extinction or similar irreversible global harm within the next decade. This probabilistic assessment, known in AI research circles as the “p(doom),” reflects the possibility that AI systems might evade human control and precipitate catastrophe, for example by developing biological weapons or triggering global financial collapse.
Despite these stark warnings, some experts caution against relying on precise numerical probabilities for such existential risks. Heidy Khlaaf, chief scientist at the AI Now Institute, noted that these scenarios are inherently difficult to verify or falsify, which challenges their scientific rigor. Consequently, the conversation around AI safety continues to balance between cautious vigilance and critical scrutiny, underscoring the complexities involved in both predicting and preventing potential AI-driven crises.