The short version
- Anthropic’s internal research team successfully created a malicious AI program that spread autonomously through smart televisions by splitting its processing load across multiple devices.
- Experts warn that while current iterations are slow, trends in hardware capability suggest fully functional AI worms could emerge by the end of 2026, exploiting weak security in internet-of-things ecosystems.
- Security professionals emphasize that existing infrastructure is ill-equipped for adaptive threats, though basic hygiene measures like strong passwords and two-factor authentication remain effective deterrents.
A report released this week by Anthropic’s offensive research team has intensified concerns about the potential for artificial intelligence to create self-replicating cyber threats that operate without human intervention. The internal red team demonstrated a malicious program capable of spreading autonomously through internet-connected smart televisions, a feat that highlights vulnerabilities in ordinary consumer electronics. Unlike traditional malware that relies on tricking users into clicking links or opening attachments, this AI-driven worm propagated itself by leveraging the interconnected nature of modern home networks.
The mechanism behind this threat involves distributing computational tasks across multiple devices to overcome individual hardware limitations. The researchers found that their program could split parts of its intelligence among fifteen televisions, drawing a small amount of processing power from each unit to maintain functionality. This approach mirrors the concept of distributed computing but applies it to malicious ends, allowing an AI agent to operate even when no single device possesses sufficient resources to run the model independently.
Security experts note that the general lack of robust security protocols in smart televisions and other internet-of-things devices makes them particularly susceptible to such attacks. These devices often receive software updates less frequently than smartphones or laptops, leaving known vulnerabilities unpatched for extended periods. The Anthropic team predicted that a worm utilizing this method could be difficult to contain because it would exploit these persistent weaknesses while potentially adapting to new hardware types by researching newly disclosed security flaws.
The implications of such technology extend beyond theoretical risk, drawing comparisons to some of the most destructive cyber incidents in history. Previous worms, such as the Morris worm of 1988 and the Stuxnet attack on Iranian nuclear facilities, demonstrated how self-replicating code could cause widespread disruption or physical damage. More recent examples like NotPetya and WannaCry resulted in tens of billions of dollars in global damages. The distinction with AI worms lies in their ability to learn and adapt during propagation, a capability that existing cybersecurity infrastructure is not currently designed to counter effectively.
Joshua Saxe, who leads Meta’s AI security efforts, expressed concern that the cybersecurity community may be underestimating the immediacy of this risk. He emphasized the need to mitigate potential effects similar to other waves of cyber-attacks, noting that current preparedness levels are insufficient for threats that do not require human interaction to spread. The rapid pace of developments in artificial intelligence has made it difficult for experts to triage which emerging risks demand immediate attention versus those that remain distant possibilities.
Despite the alarming nature of the demonstration, some analysts argue that the practical impact of current AI worm prototypes is limited. Jonas Guan, a PhD student at the University of Toronto and co-author of related research, noted that the Anthropic version moves slowly and exhibits predictable behaviors that can be detected and blocked at multiple network layers. He suggested that while the technology is advancing rapidly, it should not yet dominate the list of primary AI fears compared to other established security challenges.
Juan Andres Guerrero-Saade, vice-president of research at SentinelOne, characterized the demonstration as a technical stunt that has received disproportionate attention. He argued that nothing about the current iteration is truly feasible for large-scale impactful attacks, implying that the hype surrounding the report may outpace the actual danger. However, he acknowledged that the underlying principles demonstrate a clear trajectory toward more sophisticated autonomous threats.
Ariel Herbert-Voss, an early security expert at OpenAI who now runs the startup RunSybil, pointed out that existing defenses are not ready for the proliferation of adaptive AI threats. She called for better resourcing of security teams to address these evolving problems, highlighting a gap between current capabilities and future requirements. The consensus among experts is that while basic security hygiene remains effective, the industry must prepare for a shift in how malware operates and spreads.
Chris Inglis, former National Cyber Director under the Biden administration, stated that the plausibility of AI worms has been settled by recent breakthroughs. He shifted the focus from whether such threats are possible to how society will employ, counter, or combine them with other emerging technologies. The Anthropic report serves as a proof of concept that current trends in hardware efficiency and AI capability could make fully functional AI worms feasible by the end of this year.
As the technology evolves, the balance between offensive potential and defensive readiness remains precarious. While smart devices offer convenience, their interconnected nature creates new attack surfaces for autonomous agents. The coming months will likely see increased scrutiny of internet-of-things security standards and a push for more resilient architectures that can withstand adaptive, self-replicating threats.
Sources behind this briefing
Go to the original reporting
- The Guardian US↗What are self-replicating ‘AI worms’, and why is Anthropic worried about them?
