Anthropic releases restricted Claude Mythos model over cybersecurity concerns
Anthropic released a restricted version of its Mythos AI model, which can identify and exploit security vulnerabilities in software, after warnings from German security officials about potential misuse in cyberattacks. The public version called Fable 5 blocks responses to cybersecurity and biotechnology questions to mitigate risks, while government agencies and selected organisations retain access to the full capability through the Glasswing programme.
What happened
Anthropic released its advanced Mythos AI model in a restricted public version called Fable 5 in June 2026, after initially limiting access to preview versions available only to the US government and selected organisations. The Mythos Preview model had discovered previously unknown security vulnerabilities in widely-used software within months of its development.
German security officials had warned that criminals and state actors could use Mythos for cyberattacks on banks and energy infrastructure. To mitigate risks, Anthropic restricted Fable 5 to prevent it from answering questions about cybersecurity and biotechnology, redirecting such queries to the less capable Opus 4.8 model instead.
Reported impact
Officials and security experts expressed concerns that the Mythos model's ability to identify and exploit security vulnerabilities at unprecedented speed and precision could enable devastating cyberattacks. Mythos Preview had found security flaws in widely-used software that had remained undetected for decades.
Index analysis
Anthropic acknowledged that without protective measures, Fable 5's cybersecurity capabilities could be misused and cause severe damage. The company stated that independent experts commissioned to test Fable 5 were unable to find security flaws in the model itself.
Governance controls
Anthropic implemented safeguards including restricting Fable 5's ability to engage with cybersecurity and biotechnology topics, with sensitive queries redirected to the less capable Opus 4.8 model. The company began a programme called 'Glasswing' providing government agencies and selected companies access to Mythos models to help secure their systems against potential attacks.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.