勒索软件由 AI 自动运行，但这还不够完全

sysdig researchers documented the first known agentic ransomware attack, dubbed 'jadepuffer', where an ai agent handled the entire technical execution of a breach, from exploiting a langflow vulnerability to encrypting data and generating a ransom note. the agent acted with significant autonomy, debugging its own failed login attempts within 31 seconds and adapting to network obstacles, though it still relied on a human operator to provision infrastructure and supply the initial database credentials. this event validates concerns that ai is shifting cybercrime from a labor-intensive activity to a budget-driven one, where attackers can launch simultaneous campaigns limited only by computing costs rather than human effort. although the specific model used remains unidentified, the agent stole api keys for openai, anthropic, and others, highlighting the increased risk of data exfiltration in ai-powered attacks.