Gary Marcus

Prediction · Other global risks

“We may well see a large-scale cyberattack in which Generative AI plays an important causal role”

Gary Marcus, Substack author · 1 January 2025 · 25 AI Predictions for 2025, from Marcus on AI (Marcus on AI)

A large-scale cyberattack in which generative AI plays an important causal role may well occur in 2025.

Mostly true True in substance, wrong in a detail.

What would count as true or false

This is a strongly hedged prediction ('may well'), judged at the end of 2025. True: during 2025, a cyberattack of large scale (wide impact on many organisations or systems, or major damage) took place, and credible sources (government agencies, major security firms, or AI developers' threat reports) attributed an important causal role to generative AI, e.g. AI agents carrying out much of the intrusion, not just helping to draft phishing emails. Mostly true: such an attack was documented, but its scale was moderate or the AI role was significant but not central. Mixed: generative AI was widely documented as aiding many attacks in 2025, but no single large-scale attack with an important AI causal role was confirmed. Mostly false: there were only minor or speculative reports of AI-assisted attacks. False: 2025 threat reviews show no large-scale attack in which generative AI played an important role. Because of the hedge, judges should favour the 'mostly false' end over 'false' when such reviews are ambiguous.

Evidence

1. Disrupting an AI-orchestrated cyber espionage campaign Anthropic · 13 November 2025

   “Overall, the threat actor was able to use AI to perform 80-90% of the campaign, with human intervention required only sporadically (perhaps 4-6 critical decision points per hacking campaign).”

2. THREAT LANDSCAPE REPORT - cert.europa.eu CERT-EU (Cybersecurity Service for the Union entities) · 10 April 2026

   “A reportedly China-linked threat actor directed a jailbroken, agentic AI system against 30 entities in the technology, finance, manufacturing and government sectors, achieving autonomous intrusion in a small number of cases.”

3. Researchers question Anthropic claim that AI-assisted attack was 90% autonomous Ars Technica · 14 November 2025

   “Outside researchers weren’t convinced the discovery was the watershed moment the Anthropic posts made it out to be.”

4. Why ‘AI-Powered’ Cyber-Attacks Are Not a Serious Threat …Yet Infosecurity Magazine · 20 November 2025

   “Once again, experts questioned the claims, arguing that the absence of any indicators of compromise (IOCs) in the report means it can’t be trusted.”

Reasoning

I rated this mostly true: an attack in which AI played a central role was documented in 2025, but it reached only a small number of its 30 targets autonomously, its scale is unclear, and some of the details are disputed. Anthropic reported in November 2025 that a threat actor used AI to perform 80-90% of a cyber espionage campaign, with humans stepping in only occasionally [E1]. CERT-EU's threat landscape report describes a reportedly China-linked actor directing a jailbroken, agentic AI system against 30 entities across several sectors, achieving autonomous intrusion in a small number of cases, which matches the kind of campaign Anthropic described [E2]. Outside researchers and trade press questioned how autonomous the AI really was and whether the campaign was significant [E3, E4]. That points to a documented but disputed attack whose scale appears moderate rather than clearly large.

Cite this claim

Gary Marcus, “A large-scale cyberattack in which generative AI plays an important causal role may well occur in 2025.” (said 1 January 2025). Verdict: mostly true. Reliability API, claim c_a2c101c6e5dd, method 0.1.0+c876030c. https://epistacklabs.com/claims/c_a2c101c6e5dd