Anthropic, a prominent player in the AI field, has introduced its Claude Mythos AI model’s public variant, Fable 5, designed for identifying and addressing security vulnerabilities. Positioned as a superior option over existing AI models like OpenAI’s GPT 5.5 and Google’s Gemini 3.1 Pro, Fable is tailored for performing advanced reasoning, coding, and cybersecurity tasks.
Despite the promising features and capabilities of Fable 5, it is reported that the model is hindered by stringent guardrails established to curb misuse, such as preventing the generation of malware. These restrictions are instated with the intention of promoting safety and facilitating rapid deployment. Yet, they reportedly affect the usability of the AI by limiting its functionality under perceived risks. In circumstances where there’s a concern over security abuse, Fable 5 automatically defaults to the less capable Claude Opus 4.8. This shift is particularly challenging for cybersecurity experts who need advanced functionalities to write or review secure code since even the use of certain keywords might trigger the less functional mode.
Mythos 5, which is another variation in the series, does not carry similar restrictive safeguards and is reserved for Project Glasswing—a collaboration effort involving the U.S. government and major corporations, including tech giants like Apple, Google, and Microsoft. However, access to Mythos 5 is restricted to these members, who can leverage its full capabilities to secure essential software and systems.
Regarding cost, Fable 5 and Mythos 5 are priced at $10 and $50 per million input tokens, respectively. This pricing strategy is relatively economical compared to the pricing of the models during their preview phase. However, those wanting access to less restrictive usage of any model from Anthropic, including Fable 5 and Mythos 5, are required to undergo a Cyber Verification Program.
The concerns about the overbearing safety features stem from the dual-edged nature of AI in cybersecurity. While advanced AI tools can significantly enhance the detection and correction of system vulnerabilities, there is a substantiated fear that such tools, if misused, could aid cybercriminals in exploiting these vulnerabilities before they are patched. Illustrative of these concerns is Microsoft’s recent “Patch Tuesday” event, where it addressed almost 200 security bugs discovered through enhanced AI-assisted testing. The episode underscores a broader industry challenge where the rapid detection of vulnerabilities by AI could similarly empower nefarious activities if the technology were freely accessible.
In response to the criticisms, Anthropic has acknowledged the conservative calibration of Fable’s safeguards and noted an ongoing effort to refine the balance between usability and security. The company aims to minimize instances where innocent requests might overly limit functionality while still protecting against misuse.
The initial deployment of Fable 5 with its conservative safeguards may understandably be an avenue to cautiously monitor and adjust the system as user feedback and real-world application provide clearer insights into the necessary balance between functionality and security. This approach, while causing some dissatisfaction among cybersecurity professionals seeking to fully exploit the AI’s capabilities, highlights a cautious progression toward managing potential risks associated with powerful AI tools in cybersecurity.
This restrained approach speaks to the larger discourse within the tech industry on how to best implement and regulate powerful AI technologies. While the growth and sophistication of AI models like those from Anthropic offer remarkable potentials in areas like security, the associated risks demand careful consideration. Therefore, the ongoing dialogue among industry participants, as well as their collective experiences and adjustments, will be instrumental in shaping the tools’ efficacy and governance frameworks.
Read the full post on howtogeek.com
yyh8fi