Anthropic launched on June 9, 2026 Claude Fable 5, the primary mannequin of the Mythos class accessible for normal use. In response to the corporate, the mannequin surpasses in capability all these it has made publicly accessible so far, with frontier efficiency in software program engineering, scientific analysis and sophisticated data work.
Nonetheless, the agency launched it with mechanisms that block cybersecurity queriesbiology, chemistry and mannequin distillation, redirecting these solutions to Claude Opus 4.8.
The rationale for the blockage refers on to what Anthropic documented within the weeks main as much as the launch. By means of Undertaking Glasswing, the restricted program with which Anthropic has distributed Mythos Preview since April, the corporate and its roughly 50 companions discovered greater than 10,000 vulnerabilities of excessive or vital severity in what Anthropic describes as probably the most systemically necessary software program globally.
Analysis members embody Cloudflare, Mozilla, Microsoft, Oracle and Palo Alto Networks, in addition to banks and significant infrastructure establishments.
Cloudflare recognized 2,000 errors in vital path methods — 400 of them excessive or vital severity.with a false optimistic fee that the corporate staff thought of higher than that of human auditors. Mozilla, for its half, discovered and glued 271 vulnerabilities in Firefox 150, greater than ten occasions what it detected in Firefox 148 with Claude Opus 4.6.
The UK AI Safety Institute famous that Mythos Preview was the primary mannequin to resolve its two end-to-end cyber assault simulations autonomously.
That very same offensive energy is what Anthropic determined to limit in Fable 5. With out safeguards, the mannequin’s capabilities in areas corresponding to cybersecurity may very well be used improperly to trigger critical injury, the corporate notes in its announcement.
The corporate warns that filters are conservatively calibrated and can typically block innocent requests.
Unrestricted Mythos – Defenders Solely
Fable 5 differs from Mythos 5, its unrestricted counterpart. Mythos 5 has the identical capabilities as Fable 5, however with safeguards lifted in some areas. It’s going to initially be deployed via Undertaking Glasswing, in collaboration with the US authorities, as an replace to Mythos Preview.
Anthropic notes that it has probably the most superior cybersecurity capabilities of any mannequin on the earth. Though, in comparison with its primary opponents, Anthropic maintains one of the vital restrictive positions in the marketplace. Whereas OpenAI and Google DeepMind have progressively relaxed some filters to extend the usefulness of their fashions, Anthropic continues to prioritize safety, even at the price of limiting performance.
This “constitutional” technique has been each praised by regulators and criticized by builders searching for larger flexibility.
The context wherein Fable 5 reaches the general public is that of a documented escalation of assaults towards cryptocurrency infrastructure with AI help. In response to Maximiliano Carjuzaa, co-founder of Cash On Chain, near 100% of the DeFi assaults recorded within the earlier months concerned AI to some extent. Ledger CTO Charles Guillemet warned in April that AI is reducing the barrier to entry for attackers.
In that very same month, the trade recorded at the very least 34 hacks with losses near USD 635 million, equal to nearly 80% of the entire stolen up to now in 2026.
For the cryptocurrency ecosystem, the arrival of Fable 5 to the general public poses a selected stress: the identical filters that Anthropic considers ample for normal use haven’t been verified independently within the face of systematic evasion makes an attempt.
Fable 5 blocks cybersecurity queries, however the unrestricted model—Mythos 5—already operates within the palms of presidency and infrastructure organizations, with capabilities that Undertaking Glasswing logs exhibit as unprecedented in autonomously detecting vital vulnerabilities.
