Anthropic Warns Self-Replicating ‘AI Worms’ May Spread Across Devices Like a Virus
Key Takeaways
- •Anthropic researchers warned that self-replicating AI systems could autonomously spread across devices with basic computing power and weak security, including Android TVs, smart appliances and cameras, without human intervention.
- •Small AI models may be only months away from being capable of self-replicating on commercial IoT devices, a threat Anthropic says would be hard to contain because it would have no centralized point to shut off.
- •Glassnode co-founder Rafael Schultze-Kraft estimated that 6.26 million BTC, nearly a third of Bitcoin's supply, sits in addresses vulnerable to attack by advanced AI models or quantum computers, with 4.33 million BTC exposed due to address reuse.
- •Ethereum researcher Justin Drake urged users to migrate to fresh wallets, cautioning that AI progress could break the ECDSA signature algorithm securing both Bitcoin and Ethereum wallets sooner than previously expected.
- •Dragonfly managing partner Haseeb Qureshi dismissed the warnings as 'cryptographic doomerism,' arguing that moving tokens to fresh addresses is not a real solution to the threat.

Artificial intelligence (AI) company Anthropic has warned that self-replicating autonomous AI systems may start spreading across user devices without human intervention, behaving much like computer viruses.
In a Monday report, researchers William Zhou and Daniel Freeman of Anthropic’s Frontier Red Team — the company’s dedicated unit for probing advanced AI systems for worst-case capabilities — wrote that rogue “AI worms” may begin autonomously targeting everyday devices, including Android TVs, smart appliances, cameras and other internet-connected devices with basic computing power and weak security.
Small AI models may be “months away from being capable enough to self-replicate on commercial internet-of-things (IoT) devices,” the authors said, adding that such a threat would be difficult to contain because it would have “no centralized point to shut off.” IoT devices already number in the billions worldwide, giving any self-spreading program an enormous pool of potential hosts.
The warning follows renewed concern over the growing mathematical capabilities of AI models, which may eventually break the cryptographic signatures used in crypto wallets, threatening about one-third of the Bitcoin supply.
Almost a third of the global Bitcoin supply, or 6.26 million BTC, currently sits in addresses vulnerable to being broken by AI models or quantum computers, according to Glassnode co-founder Rafael Schultze-Kraft. In a Thursday X post, he noted that 4.33 million BTC are exposed only because of address reuse, meaning that moving those coins to fresh wallets would end the exposure. That is because a Bitcoin address is a hash of its public key: once funds are spent from a reused address, the full public key becomes visible on-chain and directly attackable, while keys that have never been broadcast stay shielded behind the hash.
A day earlier, Ethereum researcher Justin Drake warned that rapid AI advancements could break the elliptic curve digital signature algorithm (ECDSA) securing cryptocurrency wallets sooner than previously expected, and urged users to migrate to fresh wallets. ECDSA underpins Bitcoin as well, so the concern spans both of the two largest crypto networks rather than Ethereum alone.
Dragonfly managing partner Haseeb Qureshi, however, called the warnings “cryptographic doomerism”, arguing that migrating tokens to fresh addresses was not a real solution to the threat. With Anthropic pointing to a months-away horizon for self-replicating capability on IoT devices and researchers split between migration calls and pushback, the debate over how urgent the threat is remains unresolved.
Related: AI has not triggered DeFi ‘hackpocalypse,’ Dragonfly partner says