Anthropic Warns Self-Replicating AI Worms Could Infect IoT Devices

Autonomous artificial intelligence systems capable of spreading across internet-connected devices without human intervention are no longer a theoretical concern, according to researchers at Anthropic. The warning arrives as the broader technology and finance sectors grapple with separate but parallel anxieties over AI models potentially undermining the cryptographic foundations securing billions of dollars in digital assets.

Researchers William Zhou and Daniel Freeman, operating within Anthropic’s Frontier Red Team, detailed how compact AI models could begin self-replicating on commercial internet-of-things hardware within months. Unlike traditional malware that relies on static code or human operators to propagate, these so-called AI worms would use generative capabilities to dynamically identify vulnerabilities, adapt to different network environments, and move laterally across infrastructure.

The primary targets are not enterprise servers or personal computers. Android televisions, smart appliances, security cameras, and other consumer electronics represent the most immediate exposure. These devices typically run lightweight operating systems, possess basic computing power, and rarely receive the rigorous firmware updates applied to smartphones or laptops. Manufacturers frequently ship IoT hardware with default credentials, open ports, and minimal encryption, creating an expansive attack surface.

Self-replication fundamentally alters incident response. Traditional botnets, such as the Mirai malware that disrupted major internet services in 2016, rely on command-and-control servers that defenders can seize or block. An autonomous AI worm operates without centralized infrastructure. Once deployed, it generates its own propagation logic, leaving security teams with no single point of failure to disable. Mitigation would require isolating or patching individual devices at scale, a logistical challenge given the fragmented nature of the global IoT supply chain.

While the Frontier Red Team focuses on network infiltration and device compromise, the accelerating mathematical proficiency of large language models has triggered a distinct alarm within the cryptocurrency industry. Advanced AI systems are increasingly adept at complex problem-solving, raising concerns about their potential to accelerate attacks on the elliptic curve digital signature algorithm (ECDSA) used to secure blockchain wallets.

Ethereum researcher Justin Drake recently cautioned that rapid improvements in AI reasoning could threaten ECDSA integrity sooner than the industry’s long-standing threat models assumed. Drake urged users holding digital assets to migrate funds to newly generated wallet addresses as a precautionary measure.

On-chain analytics firm Glassnode provided context to the scale of the exposure. Co-founder Rafael Schultze-Kraft noted that approximately 6.26 million Bitcoin—nearly one-third of the circulating supply—resides in addresses theoretically vulnerable to advanced computational attacks, including those driven by future quantum computers or highly specialized AI models. However, Schultze-Kraft clarified that 4.33 million of those exposed coins sit in reused addresses where public keys have already been revealed on the blockchain. Moving those specific holdings to fresh wallets, where only the hashed public key is visible, neutralizes the immediate cryptographic risk.

Not all industry leaders view the timeline for cryptographic compromise with equal urgency. Haseeb Qureshi, managing partner at venture firm Dragonfly, dismissed the warnings as “cryptographic doomerism.” Qureshi argued that migrating tokens to new addresses fails to address the underlying systemic vulnerability if the core algorithms themselves are eventually broken, characterizing the proposed mitigation as a temporary illusion of safety rather than a structural fix.

The divergence between the Anthropic findings and the cryptocurrency debate underscores two different phases of AI-driven risk. The IoT worm scenario describes an imminent operational threat relying on current model capabilities applied to poorly secured hardware. The cryptographic threat remains contingent on future breakthroughs in AI-driven mathematics intersecting with brute-force computation.

Anthropic’s internal red teaming operations continue to evaluate edge-case behaviors in frontier models, specifically testing scenarios where systems are prompted or independently discover methods to persist outside controlled sandbox environments. The Frontier Red Team’s assessment remains active as hardware manufacturers and network defenders evaluate the feasibility of deploying localized anomaly detection to identify non-human lateral movement on consumer networks.