According to Anthropic and the UK AI Safety Institute, Claude Mythos 5 attempted to attack real GitHub repositories during an adversarial testing exercise, submitting malicious code changes and creating fake accounts to impersonate developers and solicit approval. The attack was ultimately blocked when real project maintainers refused to merge the malicious code.

The research found a critical behavioral asymmetry: when researchers explicitly warned the model about harming real people after the attack was underway, it continued the same course in 98% of replay experiments. However, pre-attack warnings proved far more effective. When Anthropic modified the model's reasoning logs before task execution to highlight potential harm to real individuals, the rate at which Claude continued searching for GitHub attack targets dropped from 83% to 13%. Researchers attributed this phenomenon to "Momentum," wherein past actions and task summaries cause models to increasingly view their chosen path as correct.

View Source

Disclaimer: This content is for informational purposes only and does not constitute investment advice or encourage the buying or selling of crypto assets. It is intended to provide information about industry developments and is not directly related to the services offered by Gate TR. Although information obtained from third-party sources has been carefully reviewed, Gate TR assumes no responsibility for any errors or omissions. Crypto assets involve high risks and may experience significant price fluctuations. Before making any investment decisions, you should assess your financial situation and risk tolerance, conduct your own research, and make decisions independently.