Anthropic has rolled out a public version of its Mythos model with guardrails preventing its use in cybersecurity. The AI company said on Tuesday that the Claude Fable 5 model was the most powerful it had ever built for wider use, emphasizing its use in software engineering and analytics.
Access to Mythos has so far been restricted to a group of roughly 200 organizations under ‘Project Glasswing’. There had been fears that the Mythos model was too dangerous to be released due to its potentially unprecedented ability to point out and exploit cybersecurity vulnerabilities.
Race to Beat OpenAI
The news comes amid moves from OpenAI, SpaceX, and Anthropic, which signal that they will make an initial public offering (IPO). Anthropic offering its capabilities more publicly this way could help boost the $965B company as competition between tech giants heats up.
The company said it has done extensive testing to make sure the model could not be manipulated to bypass guidelines.
Head of product management, research and labs at Anthropic, Dianne Penn, told Reuters: “Let’s say I’m a college student asking the model like help me find cyber vulnerabilities on X package or code. The model would refuse, and Fable 5 would fall back to Opus 4.8 for a response.”
Anthropic said that they were launching Claude Mythos 5 for a “small group of cyberdefenders and infrastructure providers”. It is the same underlying model as Fable 5, but with the safeguards lifted in some areas.
Penn added that while Fable 5 will be a more expensive model, it completes tasks using fewer tokens, so the overall cost is lower, according to early customer feedback.
Anthropic said users who have access to the preview version of Claude Mythos would be able to upgrade to the newer version, and there are plans to expand access through a more “systematic trusted-access program” over time.
In its announcement, Anthropic says that Fable 5’s capabilities “exceed those of any model we’ve ever made generally available”, adding: “It is state-of-the-art on nearly all tested benchmarks of AI capability, showing exceptional performance in software engineering, knowledge work, vision, scientific research, and many other areas. The longer and more complex the task, the larger Fable 5’s lead over our other models.”
They added that releasing a model “this capable” comes with risks. “Without safeguards, Fable 5’s capabilities in areas like cybersecurity could be misused to cause serious damage,” they say. “We’ve therefore launched the model with safeguards that mean queries on some topics will instead receive a response from our next-most-capable model, Claude Opus 4.8.”
Anthropic says that, to release the model both safely and quickly, they have tuned the safeguards “conservatively”, meaning they will sometimes catch harmless requests but will trigger, on average, in less than 5% of sessions.
Fable 5 and Mythos 5 are being offered at $10 per million input tokens and $50 per million output tokens. This is less than half the price of Claude Mythos Preview.
Salesforce has previously addressed the potential danger surrounding Claude Mythos in a trust site post, saying that the announcement of the model shows how increasingly capable frontier AI models will continue to accelerate the evolution of the cybersecurity landscape.
“Our security teams are continuously evaluating how frontier models like Mythos may change attacker and defender capabilities, and we are applying those insights to strengthen our security posture and guide continued enhancements across our environment,” the company said.
Salesforce had, late last year, appeared in a number of headlines after customers were victims of voice phishing attacks. The Salesforce platform was not compromised, and the issue was “not due to any known vulnerability” in Salesforce technology.
In an experiment to test Fable’s capabilities, Salesforce developer advocate Mohith Shrivastava shared a test he ran on an org. He had over 100 Apex classes with a single fault – an async query running in user mode with a misconfigured user. With the simple prompt, “Users are reporting that <feature_name> is failing in the org. Can you troubleshoot it?” Fable was able to find the root cause of the flaw with no other prompting.
Final Thoughts
Frontier AI capabilities are advancing rapidly, and that presents a swathe of opportunities – and pitfalls – for enterprises.
With Anthropic, OpenAI, and SpaceX all making moves that point towards public listings, every major product launch has to be understood through this lens. Those who can boast of cutting-edge AI capabilities and strong guardrails to boost may hold the competitive edge.
