- Joined
- Jan 20, 2026
- Messages
- 345
- Reaction score
- 2,255
Opus 4.7 now blocks textbooks, PDF with toys and Russian.
Anthropic has strengthened protective filters in Opus 4.7, but together with dangerous queries, the model began to block the usual operation. Claude Code users complain that the tool refuses to help with safe tasks: from the editing of the training laboratory on cybersecurity to reading a PDF with advertising to the toy Shrek.
Opus 4.7 was released last week after the announcement of Mythos, a model for the search and operation of vulnerabilities. Anthropic describes Mythos as a system that is too powerful for open access, and decided to use Opus 4.7 as a platform to check for stricter restrictions. The company explained that the new version automatically recognizes and blocks queries similar to prohibited or risky tasks in cybersecurity. The accumulated experience should help Anthropic prepare for a wider Hytos class models.
In practice, strict protection has hit legitimate demands. In the Claude Code repository on GitHub sharply increased the number of complaints about the classifier of the rules of acceptable use. This mechanism checks requests and decides whether they violate Anthropic’s policy. The developers write that Claude Code gives policy errors on normal tasks, and sometimes disrupts the work without an understandable explanation.
The problem did not appear suddenly, but in April the scale changed. From July to September 2025, users opened about two or three complaints per month. Among the early cases was a failure in which the memory authorization code with claude.ai caused an error API policy. In October and November, the number of similar appeals increased to about five to seven per month. In one of the messages, the developer complained that Claude 4.5 accidentally refuses to respond to regular requests.
In December, there were fewer complaints, probably due to a festive slowdown in the United States. In January, the number of appeals returned to about eight. One developer wrote that technical conversations about programming should not trigger violations of the rules, and the security filter reacts too aggressively to harmless content. In February and March, the figures remained close.
In April, the situation deteriorated sharply: the developers filed more than 30 complaints about false positives. Mistakes have affected security requests, conventional development and scientific tasks. Some users directly connect the growth of failures with the release of Opus 4.7 and the new filters that Anthropic has added to combat the dangerous use of models.
One of the most revealing cases concerned more than 40 false positives in four sessions in unrelated projects: a book on psychology, web application, infrastructure tasks and bot. And, interestingly, Claude refused to process various Russian-language requests, although the tasks did not apply to malicious activity.
In another appeal, the user complained that Opus 4.7 began to mark standard problems for computational structural biology as a violation of the rules of use. Version 4.6 dealt with the same requests. Computational structural biology studies the form and behavior of molecules using mathematical models and software tools.
Another example is the training of cybersecurity. The head of the Cybercenter and Laboratory of Applied Cyber Security at Louisiana State University, Glaudin, G. Richard III, said that Claude refused to read the laboratory on cybersecurity. The material was part of the textbook “Cybersecurity in Context” and contained simple exercises on cryptography. The author of the complaint noted that he understands the risks of using AI in attacks, but the refusal of the model to subtract the educational work for students is considered absurd, especially when subscribing more than $ 200 per month.
Not all false positives are related to cybersecurity. One developer described how Claude Code issued a policy error when trying to read a PDF with Ads to Toys Askbro Shrek. Later, the user found a fragment of the internal syntax PDF, after which the model stopped working. When deciphering, the meaningless phrase “character or for the Donkey from below” was obtained. Judging by the description, the filter did not react to the content of the document, but to a random technical sequence inside the file.
A separate failure affected security researchers, to whom Anthropic has already been given permission to work with cyber tasks. One user wrote that the exception works in Claude Chat, but does not apply when accessing Opus via the API in Claude Code. Formally, a person received the right to circumvent part of the restrictions for legitimate tasks, but the security system still blocked requests in another interface.
The developers describe the same symptom: the filter is increasingly taking normal professional work for the threat. For Claude Code, the problem is especially painful, because the tool is not used for conversations, but for the development, analysis of files, work with repositories and automation of tasks.
The increase in the number of complaints can be partly explained by the expansion of the Claude audience: the more users, the more error messages appear. But the nature of the appeals indicates not only the statistics. Claude sees a rule breach where the query relates to the legal development, training, scientific data or editing.
There is a version that the security filter is too grossly evaluates the input data. In the leaked source code of Claude Code, regular expressions were used to analyze moods, that is, search by templates in text. If the rules classifier works in a similar way and responds to individual words without a full-fledged context, false positives are almost inevitable: a term from the training laboratory, a piece of PDF syntax or phrase in Russian may look suspicious for too simple filter.
Anthropic did not respond to the request for comment. Claude Code users continue to collect bounce examples of GitHub and try to understand what words, files or formats break the work. Opus 4.7 was supposed to show how Anthropic could safely bring the public release of the Mythos level model. Instead, the first weeks turned the checking of new filters into a dispute about where safety ends and the futility of the tool begins.
Anthropic has strengthened protective filters in Opus 4.7, but together with dangerous queries, the model began to block the usual operation. Claude Code users complain that the tool refuses to help with safe tasks: from the editing of the training laboratory on cybersecurity to reading a PDF with advertising to the toy Shrek.
Opus 4.7 was released last week after the announcement of Mythos, a model for the search and operation of vulnerabilities. Anthropic describes Mythos as a system that is too powerful for open access, and decided to use Opus 4.7 as a platform to check for stricter restrictions. The company explained that the new version automatically recognizes and blocks queries similar to prohibited or risky tasks in cybersecurity. The accumulated experience should help Anthropic prepare for a wider Hytos class models.
In practice, strict protection has hit legitimate demands. In the Claude Code repository on GitHub sharply increased the number of complaints about the classifier of the rules of acceptable use. This mechanism checks requests and decides whether they violate Anthropic’s policy. The developers write that Claude Code gives policy errors on normal tasks, and sometimes disrupts the work without an understandable explanation.
The problem did not appear suddenly, but in April the scale changed. From July to September 2025, users opened about two or three complaints per month. Among the early cases was a failure in which the memory authorization code with claude.ai caused an error API policy. In October and November, the number of similar appeals increased to about five to seven per month. In one of the messages, the developer complained that Claude 4.5 accidentally refuses to respond to regular requests.
In December, there were fewer complaints, probably due to a festive slowdown in the United States. In January, the number of appeals returned to about eight. One developer wrote that technical conversations about programming should not trigger violations of the rules, and the security filter reacts too aggressively to harmless content. In February and March, the figures remained close.
In April, the situation deteriorated sharply: the developers filed more than 30 complaints about false positives. Mistakes have affected security requests, conventional development and scientific tasks. Some users directly connect the growth of failures with the release of Opus 4.7 and the new filters that Anthropic has added to combat the dangerous use of models.
One of the most revealing cases concerned more than 40 false positives in four sessions in unrelated projects: a book on psychology, web application, infrastructure tasks and bot. And, interestingly, Claude refused to process various Russian-language requests, although the tasks did not apply to malicious activity.
In another appeal, the user complained that Opus 4.7 began to mark standard problems for computational structural biology as a violation of the rules of use. Version 4.6 dealt with the same requests. Computational structural biology studies the form and behavior of molecules using mathematical models and software tools.
Another example is the training of cybersecurity. The head of the Cybercenter and Laboratory of Applied Cyber Security at Louisiana State University, Glaudin, G. Richard III, said that Claude refused to read the laboratory on cybersecurity. The material was part of the textbook “Cybersecurity in Context” and contained simple exercises on cryptography. The author of the complaint noted that he understands the risks of using AI in attacks, but the refusal of the model to subtract the educational work for students is considered absurd, especially when subscribing more than $ 200 per month.
Not all false positives are related to cybersecurity. One developer described how Claude Code issued a policy error when trying to read a PDF with Ads to Toys Askbro Shrek. Later, the user found a fragment of the internal syntax PDF, after which the model stopped working. When deciphering, the meaningless phrase “character or for the Donkey from below” was obtained. Judging by the description, the filter did not react to the content of the document, but to a random technical sequence inside the file.
A separate failure affected security researchers, to whom Anthropic has already been given permission to work with cyber tasks. One user wrote that the exception works in Claude Chat, but does not apply when accessing Opus via the API in Claude Code. Formally, a person received the right to circumvent part of the restrictions for legitimate tasks, but the security system still blocked requests in another interface.
The developers describe the same symptom: the filter is increasingly taking normal professional work for the threat. For Claude Code, the problem is especially painful, because the tool is not used for conversations, but for the development, analysis of files, work with repositories and automation of tasks.
The increase in the number of complaints can be partly explained by the expansion of the Claude audience: the more users, the more error messages appear. But the nature of the appeals indicates not only the statistics. Claude sees a rule breach where the query relates to the legal development, training, scientific data or editing.
There is a version that the security filter is too grossly evaluates the input data. In the leaked source code of Claude Code, regular expressions were used to analyze moods, that is, search by templates in text. If the rules classifier works in a similar way and responds to individual words without a full-fledged context, false positives are almost inevitable: a term from the training laboratory, a piece of PDF syntax or phrase in Russian may look suspicious for too simple filter.
Anthropic did not respond to the request for comment. Claude Code users continue to collect bounce examples of GitHub and try to understand what words, files or formats break the work. Opus 4.7 was supposed to show how Anthropic could safely bring the public release of the Mythos level model. Instead, the first weeks turned the checking of new filters into a dispute about where safety ends and the futility of the tool begins.
