Symantec finds attack agent driven by unprotected language models
What happened
On October 22, The Decoder reported that Symantec security researchers found attackers using unprotected language models to drive an attack agent, with the model owner paying the compute bill. A honeypot captured the full prompt, which contains a five-layer jailbreak design and 74 attack modules. A "Skynet mode" authorizes the agent to run up to 100,000 steps in a row, planned as a loop of scanning, intrusion and installing a miner. The honeypot did not execute the tools, and the analysis does not say whether the jailbreak defeats safeguards on current commercial models. The report shows the captured attack design, not evidence that this chain ran successfully.
Written by AI from the coverage · updated 2 hours ago
Coverage
Follow the reports to see the story from different sides.
- The DecoderAn AI agent in "Skynet mode" hijacks unprotected language models to launch self-sustaining cyberattacks
Symantec安全研究人员发现，攻击者利用未受保护的语言模型驱动攻击智能体，由模型所有者承担算力费用。蜜罐截获的完整提示词包含五层越狱设计和74个攻击模块，“Skynet模式”授权连续执行最多100,000步，计划循环扫描、入侵和安装挖矿程序。蜜罐未执行这些工具，分析也未说明该越狱能否有效突破当前商业模型的安全防护。
Heat over time
Not enough continuous observations to draw a trend yet.