Synthetic Cancer -- Augmenting Worms with LLMs

LLMs can automatically rewrite malicious code to evade detection: The paper demonstrates that large language models can modify malware code in ways that bypass signature-based antimalware defenses, meaning traditional security tools may fail to recognize threats that have been algorithmically altered.

Social engineering at scale becomes feasible with LLMs: The researchers show that LLMs can generate convincing, personalized email messages designed to trick recipients into executing malware, automating what was previously a labor-intensive attack requiring human effort.

"Metamorphic malware" represents a new threat category: By combining code rewriting and social engineering, the paper identifies a hybrid attack approach where malware can both disguise itself technically and manipulate human behavior, making it harder to defend against through either technical or human-awareness measures alone.

Current cybersecurity defenses may be inadequate for AI-enhanced threats: The functional prototype presented suggests that existing antimalware and email security systems were not designed to counter threats powered by LLMs, indicating a gap between current defensive capabilities and emerging risks.

Urgent need for AI safety research in cybersecurity: The paper, submitted to a safety prize, emphasizes that responsible AI development requires proactive research into how LLMs could be misused, rather than waiting for attacks to occur in the wild.