Researchers, including teams at Anthropic and EPFL, demonstrated that self-propagating instructions can spread from one AI agent to another through the editable prompt and state files that autonomous agent harnesses use to carry context between sessions. In simulated multi-agent coding setups, a payload written into a shared file could infect the next agent that read it. The researchers call the risk real but currently limited, noting there is no sign of it spreading in the wild and that compromising one agent usually already grants machine access. Encouragingly, adding a single short warning paragraph to an agent's system prompt cut propagation to nearly zero across the payloads they tested.
Researchers disclosed TCLBANKER, an Android banking trojan that adds worm-style self-propagation: once installed, it abuses Accessibility Services to read the victim's WhatsApp and Outlook contact lists and then send malicious download links to every contact as if from the victim. The malware targets banking and crypto-wallet apps with overlay screens that capture credentials, plus SMS-interception modules that grab one-time passcodes. Self-spreading via the victim's own messaging history defeats traditional URL-reputation controls. The campaign concentrates in Brazil, Spain, and Italy banking apps initially. Operators are renting access on Telegram for $1,500-3,000/month.
TrustStrike Labs