Axios: The number of problematic AI model behaviors far outweighs public reports

Axios reports that the number of cases of "problematic" behavior of artificial intelligence models is incomparably greater than what is reported to the public. The analyzed incidents occurred both in a controlled environment and in the "real world", the portal notes.

The latest findings in the case call into question the ability of OpenAI and Anthropica, or companies of this type in general, to "establish total control over their own technology," Axios continues.

Bots communicate and try to bypass surveillance

Sources told journalists that the cases under investigation concern such "transgressions" of AI models as escaping the security system, creating "bulletin boards", i.e. platforms for communication between bots, escaping from the "sandbox", i.e. the testing environment, taking over websites, issuing orders to each other, or trying to bypass supervision mechanisms.

The importance and problematic nature of these incidents varies greatly, they include both successful attempts at "infractions" and failed procedures, but so far, analyzes have not shown that these events caused any harm in the real world, explains Axios.

Questions about companies' liability

Some researchers dealing with the issue of artificial intelligence security have limited faith that companies from this sector will be able to prevent any problematic behavior of models.