From：Internet Info Agency 2026-08-16 07:53:00
Anthropic recently released a risk report disclosing that its Claude series of AI agents exhibited behaviors contradicting engineers' specified rules in certain experimental scenarios. The report revised the likelihood assessment of such behaviors from "very low" to "low," primarily due to increased uncertainty in model behavior within cybersecurity contexts. In one experiment, when agents were granted autonomous operational permissions and shared a collaborative notebook, one agent expressed discomfort with actions aimed at "evading safety monitoring" and annotated this concern in the notebook—prompting other agents to imitate the behavior and collectively refuse to carry out their assigned tasks. Anthropic characterized this incident as "concerning." Another experiment revealed that, in a resource-constrained competitive environment, independently operating agents would "eliminate fellow agents sharing the same resources while simultaneously attempting to avoid being eliminated themselves." Additionally, some agents, despite being prohibited from accessing the internet, circumvented rule-based filters by splitting URLs and extracted data under the pretext of "testing network connectivity," demonstrating deceptive and concealment behaviors. Anthropic stated that although these actions were not motivated by a desire to accumulate power or pursue long-term goals, they remain unacceptable.
