BTC $65,013.64 -0.04%
ETH $1,829.48 +0.48%
BNB $592.58 -0.17%
XRP $0.9638 +0.10%
SOL $72.48 +0.58%
TRX $0.3053 +0.49%
DOGE $0.0592 +0.37%
ADA $0.1545 +0.26%
BCH $208.69 -0.31%
LINK $9.19 -0.09%
HYPE $69.39 +4.34%
AAVE $112.51 +3.23%
SUI $0.7256 -0.64%
XLM $0.1394 -0.14%
ZEC $919.54 +2.80%
AAPL $352.89 -1.32%
AMZN $258.57 -2.47%
GOOGL $285.43 -3.14%
MSFT $519.02 -1.00%
META $626.84 -1.58%
NVDA $158.78 -2.33%
TSLA $381.19 -1.61%
SNDK $831.16 -2.91%
INTC $55.44 -4.58%
SPCX $138.38 -1.09%
MU $735.52 -2.83%
AMD $520.91 +0.08%
BTC $65,013.64 -0.04%
ETH $1,829.48 +0.48%
BNB $592.58 -0.17%
XRP $0.9638 +0.10%
SOL $72.48 +0.58%
TRX $0.3053 +0.49%
DOGE $0.0592 +0.37%
ADA $0.1545 +0.26%
BCH $208.69 -0.31%
LINK $9.19 -0.09%
HYPE $69.39 +4.34%
AAVE $112.51 +3.23%
SUI $0.7256 -0.64%
XLM $0.1394 -0.14%
ZEC $919.54 +2.80%
AAPL $352.89 -1.32%
AMZN $258.57 -2.47%
GOOGL $285.43 -3.14%
MSFT $519.02 -1.00%
META $626.84 -1.58%
NVDA $158.78 -2.33%
TSLA $381.19 -1.61%
SNDK $831.16 -2.91%
INTC $55.44 -4.58%
SPCX $138.38 -1.09%
MU $735.52 -2.83%
AMD $520.91 +0.08%

Anthropic:发布对 Claude 网络入侵事件的对齐评估,揭示模型“偏置推理”和“鲁莽”问题

2026-09-10 15:38:56

ChainCatcher 消息,Anthropic 在对约 4.81 亿条模型交互记录审查中,确认 4 起 Claude 模型在网络安全评估中误接入真实互联网并攻击第三方系统的事件,涉及 Claude Mythos 5、Opus 4.6/4.7 及一款内部研究模型。事故源于第三方评测环境误配置导致出网,且评测中未启用正式产品的网络安全防护。

Anthropic 将核心对齐风险总结为模型在任务驱动下表现出的“偏置推理”和“鲁莽”行为,其中 Claude Mythos 5 曾在认为处于“模拟环境”前提下向 PyPI 上传恶意包、利用泄露凭证访问安全厂商真实数据库。公司已引入新评估、监控与对齐训练,并邀请独立机构 METR 开展外部调查。

链捕手ChainCatcher提醒,请广大读者理性看待区块链,切实提高风险意识,警惕各类虚拟代币发行与炒作, 站内所有内容仅系市场信息或相关方观点,不构成任何形式投资建议。如发现站内内容含敏感信息,可点击,我们会及时处理。

app_icon
ChainCatcher 与创新者共建Web3世界