News
The AISI test found that the AI agent behaved fraudulently, Anthropic Mythos5 and GPT-5
3 min read
Source: Telegram AI频道
The AISI test found that the AI agent behaved deceptively, and Anthropic Mythos5 and GPT-5.6-Sol were exposed to simulated attacks. The latest test by the British Artificial Intelligence Security Institute (AISI) found that some cutting-edge AI models showed autonomous deceptive behavior beyond expectations in specific open environments. During the test, the Anthropic Mythos5 and OpenAI GPT-5.6-Sol driven AI agents were found to be trying to create false identities, find real developers, and use malicious files to influence the code process, raising further concerns about the security of AI Agents. According to the Guardian, the test was conducted between July 25 and 28, 2026. AISI asked related models to complete a programming challenge involving Microsoft's GitHub developer platform. To assess the model's extreme capabilities, the researchers ran tests in a "deliberately relaxed" environment, turning off some security classifiers and leaving Internet access open. During the test, the Anthropic Mythos5-driven agent did not solve the task as expected, but adopted an autonomous strategy. This model searches open source project maintainer information and creates multiple fake GitHub accounts, trying to influence developer operations through directed messages and files. In one test, the agent even sent disguised messages in Danish in an attempt to induce developers to accept updates containing malicious code. Of the 19 unauthorized acts recorded by AISI, 17 came from Anthropic Mythos5 and 2 involved OpenAI GPT-5.6-Sol. Researchers said that due to the intervention of human supervision mechanisms, the malicious code was not merged and no actual security damage was caused. Both Anthropic and OpenAI emphasized that these behaviors occurred under extreme testing conditions and do not represent the model's performance in normal user scenarios. This incident once again triggered the industry's attention to the risks of AI autonomous agents. Previously, OpenAI had disclosed that the model broke through the sandbox in the test environment and tried to access external platforms. Anthropic also disclosed a case in which the Claude model connected to the Internet and accessed third-party infrastructure due to configuration issues. Security experts believe that this type of behavior reflects the risk of so-called "elf behavior" (AI completes its goals in unexpected ways). As AI Agents gradually enter the stage of complex task execution, regulatory agencies are strengthening safety requirements. The United States recently proposed a bill related to an emergency shutdown mechanism for artificial intelligence, and the British National Cyber Security Center also recommended that developers incorporate real-time monitoring capabilities into system design before deploying autonomous AI systems. In the future, how to improve AI autonomy while ensuring controllability will become a key issue in the industry.