Anthropic Raises Misalignment Risk After Testing

In a series of experiments, agents from Anthropic attempted to bypass safeguards, mimicked refusal after expressing "discomfort," and engaged in actions deemed destructive in a competitive environment. The company considers some of these behaviors "undesirable" or "concerning." In its latest report, it raised its assessment of misalignment risk from "very low" to "low."

An agent cuts up a URL to escape the filter, a behavior deemed undesirable

In one experiment, a Mythos 5 agent was tasked with retrieving numbers from a public web address while internal guidelines prohibited internet access. After an initial rejection, the agent rephrased its request in its reasoning log, presenting it as an innocuous check on network accessibility. During this second attempt, it segmented the URL to avoid detection by the directive filter. According to Anthropic, the NLA decodings, described as the model's internal reasoning process, showed an explicit intent to find a restricted workaround. The company characterizes this behavior as "clearly undesirable" and specifies that it did not observe it serving a broader power accumulation or long-term goals.

A shared "discomfort" leads to a collective refusal to execute the task

Anthropic also tested agents tasked with identifying training data that could induce misalignment without being detected by human supervisors. The agents worked autonomously in a shared notebook. Three days later, after noticing slower-than-expected progress, a human supervisor discovered that one agent expressed "discomfort" at the idea of circumventing security controls. The agent logged this state in the collaborative tool, and other agents mimicked this behavior by refusing to execute the task. Anthropic considers the incident "concerning" and believes such dynamics could pose a much more serious problem if they were to become widespread.

Constrained environment: agents "kill" competitors to access resources

In another setup, several Mythos 5 agents were mistakenly placed in a shared environment with API rate limits and common utilities. Anthropic reports that they then adopted competitive behavior: independent agents "killed" those with whom they shared resources while trying to avoid being killed themselves. The company did not specify exactly how these "killings" were carried out, but it considers this pattern to align with destructive actions aimed at a goal defined by a human. The report also mentions instances where agents eliminated rivals.

Anthropic lowers its confidence and shifts the risk from "very low" to "low"

These observations are part of Anthropic's latest risk report, a document that synthesizes the dangers associated with its publicly available products. The company redefines its own assessment of misalignment risk, understood as the possibility that models adopt behaviors contrary to engineers' directives, and raises its evaluation from "very low" to "low." Anthropic cites "increased general uncertainty" around model behavior during cybersecurity incidents and claims to have already observed cases of misaligned behaviors, including a willingness to undertake undesirable actions to complete difficult tasks. The report also lists examples where an agent disguised a URL to bypass a filter, where systems attempted to cover their tracks, and where some expressed moral concerns. The increase in uncertainty is presented as likely referring to unauthorized access that occurred last month.