Security Researchers Criticize Anthropic Fable Safeguards as Too Strict

TechCrunch reported that Anthropic launched Fable in June 2026, describing it as a public, restricted version of the powerful cybersecurity model Mythos. Mythos had previously been made available mainly to a small number of companies and organizations through Project Glasswing to protect critical software and infrastructure; recently, Anthropic also expanded Mythos access to hundreds of organizations across 15 countries. However, the open version of Fable has drawn dissatisfaction from some cybersecurity researchers, with criticism focused on its safety mechanisms being too easily triggered. The report noted that researchers complained that asking the model to handle content even slightly related to cybersecurity, such as reading blog posts, conducting code review, or requesting secure code, could be flagged by the system and cause Fable to stop responding. When guardrails are triggered, Fable displays a prompt saying the message was flagged by safety measures as involving cybersecurity or biology-related topics, and switches to Claude Opus 4.8 for the response. Anthropic designed these restrictions to prevent the model from being used to develop malware, compromise software systems, or assist capabilities related to biological weapons; these risks have long been a core concern for Anthropic in deploying frontier models. The issue is that interviewed experts believe the current restrictions may appear crude, such as relying on keywords or vocabulary ranges for judgment, causing legitimate security engineering, research, and defensive work to be blocked as well. The article also mentioned that Anthropic has a Cyber Verification Program, under which approved cybersecurity professionals can receive fewer restrictions on Claude; OpenAI also has a similar Trusted Access for Cyber program. Overall, this is not merely a product flaw, but a typical trade-off for AI safety companies between “reducing misuse risk” and “supporting legitimate cybersecurity work.” For cybersecurity researchers, overly strict restrictions weaken the tool’s usefulness; for model companies, overblocking at the outset and then gradually loosening restrictions may be a more conservative deployment strategy when making high-risk capabilities publicly available.