Anthropic officially released its latest model, Fable, to the public on Tuesday, positioning it as a "public, restricted version" of its internal high-end cybersecurity model Mythos, but the product quickly sparked controversy in cybersecurity circles. Many security researchers and practitioners complained on social platforms and communities that Fable’s built-in security guardrails were too strict and could hardly be used for any actual network security-related work.
According to feedback from researchers, Fable rejects "any requests even remotely related to cybersecurity," including even seemingly innocuous tasks such as helping to read a blog post. According to researcher Valentina “Chompie” Palmiotti, who now works at IBM X-Force, Fable simply terminates the conversation and indicates that its security mechanisms have flagged the message as involving cybersecurity or biological topics. These guardrails are designed to prevent models from being used to develop malware, attack or damage software systems, and also to limit their misuse in the biological field to assist in the development of biological weapons.
When Anthropic launched Mythos in April of this year, it chose to open it to only a small number of enterprises and institutions through a program called "Project Glasswing" with the intention of using this model to help protect critical software and infrastructure. Last week, Anthropic announced that it would expand the use of Mythos to hundreds of organizations in 15 countries, further promoting the implementation of this type of high-capability security model in key industries. However, after Fable was opened to the public, its "downgraded version" security strategy was strongly questioned among professional users. Many people believed that there was a significant gap between the actual experience and the official propaganda.
Matt Suiche, a longtime cybersecurity veteran, told TechCrunch that Fable was very abrupt in determining whether a request was cybersecurity-related. For example, he said that if a user asks for "writing secure code," Fable will tend to regard it as network security work rather than software engineering best practice guidance, thus directly triggering the downgrade mechanism. Once the guardrail is triggered, Fable will automatically fall back to the less capable Claude Opus 4.8 to continue the conversation. Suiche believes that Fable's judgment logic seems to be highly dependent on keywords. "As long as words fall in the semantic field of 'network security', they can easily be intercepted by the security system."
Despite this, Suiche also expresses a certain understanding of the stringent settings at the current stage, believing that at this early stage, manufacturers impose more conservative safety thresholds on models and are more secure in risk control. He expects these guardrails will be continually refined and fine-tuned as Anthropic deepens its collaboration with a new generation of cybersecurity companies. In his view, it is a more acceptable path to "block more" first and then gradually relax restrictions than to relax too much at the beginning, causing the potential risk of abuse to get out of control.
Fable is not alone in his dissatisfaction. Another researcher complained on the social platform that "even requesting it for code review will trigger the safety guardrail." Some users shared their experiences in Reddit's Claude-related community, saying that Fable "almost all refuses" requests for security audits, vulnerability analyses, etc., seriously affecting its usefulness in professional environments. As of press time, Anthropic has not publicly responded to the feedback.
In addition to the automatic guardrail mechanism within the model, Anthropic has also established an additional admissions process for cybersecurity practitioners - the "Cyber Verification Program". Only users who pass the program can use Claude for network security work under less restrictive conditions. Similarly, OpenAI has launched a project called "Trusted Access for Cyber" to open up more model capabilities for compliant cybersecurity practices. These practices reflect that while cutting-edge model companies are promoting AI-empowered network security, they are still trying to balance the release of capabilities and the risk of abuse through the dual means of review systems and technical guardrails.