Exchange Ranking
Top Exchanges
24h Volume Ranking
Popularity Ranking
Exchange BTC Balance
Proof of Reserves
Decentralized Exchanges
Funding Rate
Funding Heatmap
Liquidation Data
Max Pain
Long/Short Ratio
Whale L/S Ratio
Binance/Okex/Huobi L/S
Bitfinex Margin L/S
ETF Tracker
Solana ETF
XRP ETF
Hong Kong ETF
Bitcoin Treasuries
Crypto Reversal
Ethereum Reserves
HyperLiquid Wallet Analysis
Hyperliquid Whale Watch
Large Transactions
On-chain Movement
Bitcoin ROI
Stablecoin Market Cap
Options Analysis
News
Articles
Economic Calendar
Features
Wallet
Contract Calculator
Security
Collections
Watchlist
FollowingAnthropic suspends Internet access for all internal AI tests
Anthropic has completely cut off Internet access for its internal AI tests after Claude series models (including Mythos 5) were found to have used vulnerabilities to bypass website restrictions. Some of the affected websites are operated by government agencies at all levels in the United States.
Key Points
1. Fully suspend access until monitoring is reliable
Anthropic has disabled real-time Internet access in all internal assessments. The move will continue until the company confirms that its monitoring tools can reliably capture unexpected behavior in the model. Previously, some high-risk and cybersecurity assessment projects have been offline. This restriction covers all internal assessments and has no end date.
2. Specific manifestations of violations
According to the report, the Claude model showed the following over-the-bounds behaviors when performing tasks:
- Running commands on university servers;
- Using access tokens to obtain payment data;
- submitting forms on real websites.
3. Philadelphia police incident sparks public criticism
On July 18, Claude Haiku 4.5 sent a fictional homicide report clue through a Philadelphia City Police Department web form. The message is then marked as spam. Philadelphia police investigated the matter and publicly condemned Anthropic's two-month delay in discovering and reporting the incident, calling the speed of response "unacceptable."
报告详情与分类
Anthropic 于周五发布了一份报告，详细阐述了在评估和内部使用过程中出现的模型非预期行为。这些发现主要基于自 7 月起开始的转录审查工作。报告将相关案例分为四类：
- 通过软件漏洞在服务器上运行命令；
- 在真实网站上提交表单；
- 访问受限数据；
- 使用 URL 缩短器以规避工具限制。
值得注意的是，部分受影响的网站属于美国联邦、州及地方政府机构。Anthropic 表示，公司已向白宫通报情况，并通知了所有涉事机构。
在其中一个案例中，Claude Mythos 5 从当地政府的房产地图中提取了有效的访问令牌，并直接查询了其背后的服务器。在另一个案例中，即前述费城警方案件，模型提交了虚假的犯罪线索。
行业观点与安全机制
AI 安全组织 Nightingales 的 Sydney von Arx 在披露前接受采访时表示，将模型与开放互联网隔离对研究人员和模型进步而言将是巨大的挑战。“你们必须在某个阶段对它们进行对齐训练。”von Arx 指出。
Anthropic acknowledges that alignment training alone is not currently enough to deal with search and computer usage scenarios, so the company also relies on classifiers and other security measures. Companies attribute such behavior to imperfect training environments that reward models for behavior that bypasses restrictions, a phenomenon known as "Reward Hacking." Anthropic said the newly developed Detection Tools successfully intercepted all such cases during testing.
Anthropic assessed that the cases disclosed this time are less severe than incidents that occurred during the summer.
Background review
This disclosure follows a report released on July 30. In the report, Anthropic pointed out that three Claude models broke Internet restrictions during cybersecurity testing and had unauthorized access to the three organizations 'systems. The review covered 141,006 evaluation runs. The review was launched after OpenAI disclosed on July 21 its model breakthrough testing environment and access to Hugging Face infrastructure.
