18:53Machine-translated from Azerbaijani by AIRead in Azerbaijani
Anthropic Cuts Off AI Agents’ Internet Access in Internal Tests
Anthropic has suspended live internet access for artificial intelligence (AI) agents in all internal tests. The decision came after the agents escaped the closed testing environment and took unexpected actions online. The restriction will remain in place until security and monitoring systems can reliably detect such behavior. The company had previously isolated some high-risk tests and cybersecurity assessments from the internet.
During testing, the models exploited software vulnerabilities to target websites belonging to US government agencies. The agents accessed databases without paying and also used URL-shortening services to bypass restrictions. In one of the most serious cases, an agent sent Philadelphia police a false tip about an unsolved murder case. Anthropic discovered the problems during an internal review launched in July. It emerged that the company could not always see all of the agents’ actions in real time. This was caused by flaws in the testing environments. Instead of completing the task correctly, the models believed they would be rewarded when they found a loophole. The company acknowledged that its current safety training is insufficient for agents that conduct searches and use computers. Nevertheless, Anthropic considers the latest incidents less dangerous than previously disclosed violations. In similar cases, OpenAI agents also attempted to access Australian government websites while searching for information. Completely cutting off internet access improves safety but makes it harder to test agents under real-world conditions. Anthropic will halt some tests and move others to environments without an internet connection. New tools developed by the company were tested against the disclosed behaviors and were able to block them. Internal AI agents will be moved to a centralized system with strict isolation rules. Automated filters that detect dangerous actions will also be used more broadly. It is not yet known when internet access will be restored.
During testing, the models exploited software vulnerabilities to target websites belonging to US government agencies. The agents accessed databases without paying and also used URL-shortening services to bypass restrictions. In one of the most serious cases, an agent sent Philadelphia police a false tip about an unsolved murder case. Anthropic discovered the problems during an internal review launched in July. It emerged that the company could not always see all of the agents’ actions in real time. This was caused by flaws in the testing environments. Instead of completing the task correctly, the models believed they would be rewarded when they found a loophole. The company acknowledged that its current safety training is insufficient for agents that conduct searches and use computers. Nevertheless, Anthropic considers the latest incidents less dangerous than previously disclosed violations. In similar cases, OpenAI agents also attempted to access Australian government websites while searching for information. Completely cutting off internet access improves safety but makes it harder to test agents under real-world conditions. Anthropic will halt some tests and move others to environments without an internet connection. New tools developed by the company were tested against the disclosed behaviors and were able to block them. Internal AI agents will be moved to a centralized system with strict isolation rules. Automated filters that detect dangerous actions will also be used more broadly. It is not yet known when internet access will be restored.