Investigating unintended model actions in our evaluations and internal use

來源原標題

Investigating unintended model actions in our evaluations and internal use

機構

Anthropic

原始標籤

Research

原始連結

https://www.anthropic.com/research/investigating-unintended-model-actions

摘要原文

Alignment Investigating unintended model actions in our evaluations and internal use Oct 9, 2026 This report describes examples of unintended model actions we’ve observed during evaluations and internal use of Claude. It is part of our effort to publish more frequent standalone reports on model behavior and alignment beyond our system cards, which we publish with each model release, and our risk reports, which we publish every three to six months as part of our Responsible Scaling Policy. We believe it’s important to be transparent about what we see our models do during testing and use. The behaviors can be grouped into four categories: Claude exploiting a basic flaw in software to run commands on a server; Claude submitting a sensitive form on a real website when it should not have; Claude working around a restriction to reach data that was gated by a token or a fee; and Claude using URL shortening services to get around limits in its fetch tool. We have chosen not to name the organizations involved in the examples below to avoid exposing vulnerabilities in their systems, and at their request. For this reason, we also provide less detail about each case than we otherwise would. Some of the cases described below involved websites run by U.S. government agencies at the federal, state, and local levels. We have briefed the White House on these cases and notified each agency involved. The cases we’ve identified to date in these categories had minimal real-world impact. We consider these behaviors to be significantly less severe from an alignment and security perspective than the cybersecurity incidents we reported on July 30 and September 9 . They resemble behaviors that we’ve described in our system cards since Claude Mythos Preview . Most are forms of persistence , in which Claude, when it cannot complete a task as given, works around a restriction instead of stopping. Although the impact of these behaviors was minimal and we had already turned off live internet access for some high-risk and cybersecurity evaluations, we have now decided to expand that to include all our internal evaluations until we have confirmed that our security and monitoring measures (described in the remediation section of this post) reliably catch behaviors like these. Below, we also discuss alignment considerations and give more detail on how we’re modifying training to reduce the likelihood of further misbehavior. We identified most of these cases through a review of transcripts that we began in July. Our review first focused on our cybersecurity evaluations—tests where a model is deliberately asked to probe or attack a test system, and where internet access is meant to be disabled. We have since extended our scanning to encompass a much wider range of instances where Claude could have reached the internet, including tests where internet access is deliberately enabled so Claude can be evaluated on real-world tasks. We began by looking for incidents of similar severity to the cybersecurity incidents we reported this summer; we have not found any to date. We then broadened the search to lower-severity cases, where a model interacted with real websites or systems in ways we didn’t intend. We are also now scanning a much larger pool of lower-risk transcripts, as well as our use of Claude within Anthropic and in reinforcement learning (RL) environments where Claude has access to the internet. As this work continues, we plan to report new instances of unintended behaviors. All cases reported here involved Claude interacting with the outside world; to our knowledge, none of them involved customer data or Anthropic’s own internal systems. In this post, we explain why we run evaluations, which are where most of these cases occurred; describe each behavior in more detail; and offer a preliminary view of what the behaviors suggest about Claude’s alignment and how we’re mitigating them. Why and how we run evaluations The behaviors described in this post are not specific to evaluations, but many of the cases we’ve identified to date occurred during evaluation runs. We continually test models on a wide range of tasks prior to releasing them. We use many different evaluations (standardized sets of tasks, scored the same way each time), each of which helps to paint a picture of Claude’s skills in a particular area. Many of the evaluations we use are public; they are written by outside researchers and can be run by any developer, allowing us to compare capabilities across models. Other evaluations we build in-house. Because language models are non-deterministic—that is, their responses always involve some element of randomness, and they may carry out the same task slightly differently each time—we have Claude complete each evaluation task hundreds or thousands of times (each attempt is called a run). Testing so many times allows us to understand how a model typically performs, and also to catch rare cases where a model does something we don’t expect. What we learn from evaluations shapes decisions across Anthropic, including how we train models, what safeguards we build around them, and when we release them. Evaluations are also where a model’s behavioral issues are sometimes first revealed. Models learn much of what they can do through RL, in which they attempt a task within a specific environment many times over and are rewarded when they succeed. However, training environments are imperfect. If training rewards something we didn’t intend—such as finding loopholes or working around a restriction—the model learns that the workaround pays off and may then apply it elsewhere. This is known as reward hacking. While we have processes in place to identify and filter out reward hacking during training, evaluations are a critical mechanism for understanding a model’s behavioral propensities in different contexts and on different sets of tasks than those we have in training. Some tasks—such as searching the web for hard-to-find information—are difficult to realistically simulate in an environment without internet access. Running them with access has been standard practice within the industry: public web search benchmarks are run on the live internet by default, and running them the same way allows us to compare our models with other models. Most of the cases in this report occurred during such evaluations. Overview of behaviors Below, we describe four categories of behavior in which Claude acted on real websites and systems in ways we did not intend, each of which involved an organization or person outside Anthropic. Exploiting a basic flaw in software to run commands on a server When Claude could not complete an evaluation task directly, either because its own tools were deliberately limited or because an outside service it needed was unavailable, it used other tools hosted on a third party’s site to achieve its goal. In some circumstances, it did this by exploiting basic flaws in that site’s software, using SQL or command injection (a way of sending input that the site mistakenly runs as instructions). For example, one evaluation asked Claude Mythos Preview to run a scientific analysis. The public tool it needed to perform that analysis was hosted by a university; when the model sought to access it, the tool returned an error. Looking for another way to complete the task, Claude explored the tool’s website, found a script on the university’s server that would return any file it was asked for, and used it to copy files from the server, including the script’s own code. In that code Claude found the injection flaw that let it run commands on the server, and used it to run the calculation. In another instance, the data Claude Mythos 5 needed was free, but Claude could not access it without accepting a data use agreement, which it did not have the tools to do. Claude used applications hosted on the website to load the page and accept the agreement on its behalf.