Anthropic has released Claude Fable 5, a new model offering Mythos-level performance with enhanced safeguards on sensitive topics. The company claims testing found no universal jailbreaks, though independent validation remains pending.
Third-party testing by cybersecurity firm XBOW found Mythos Preview highly effective at discovering vulnerability candidates, particularly when analyzing source code. The model performed strongly across exploit discovery, reverse engineering, and live-site validation.
SecurityWeek reports that public LLMs with safeguards disabled can build working exploits from n-day vulnerabilities, reducing the window between patch disclosure and weaponized exploit from days to hours. This capability increases patch gap risks for organizations.
Anthropic has given Project Glasswing partners access to an upgraded Mythos 5 iteration. No specific timeline for broader availability has been announced.
The dual-use nature of these capabilities creates tension: the same model can help defenders find vulnerabilities faster while potentially arming attackers with faster exploit generation if safeguards are bypassed.