Anthropic Releases Claude Model with Cybersecurity Restrictions
Anthropic released a restricted version of its Mathos AI model as Claude Sonnet 5, prohibiting use in cybersecurity and biotechnology after an earlier version discovered decades-old vulnerabilities in common software. The company implemented safeguards and provided select government agencies and companies early access through a program called Glass Wing to help secure their systems.
What happened
Anthropic released a public version of its Mathos model called Claude Sonnet 5, with restrictions preventing use for cybersecurity and biotechnology applications.
Months earlier, the Mathos Preview model discovered security vulnerabilities that had remained hidden for decades in widely-used software, raising concerns about AI being used to find vulnerabilities and conduct destructive cyberattacks.
Reported impact
The discovery of hidden security vulnerabilities in established software created concerns about potential misuse of AI to identify and exploit system flaws.
Index analysis
Anthropic implemented safeguards and protective controls against potential dangerous uses of the Claude Sonnet 5 model and subjected them to intensive testing.
Governance controls
Selected government agencies and companies were granted access to Mathos Preview to secure their systems against potential attacks through a program called Glass Wing.
The public Claude Sonnet 5 release includes restrictions on cybersecurity and biotechnology applications to mitigate misuse risks.
Classification
Sources and evidence
Sources not yet published
This record was researched and written by the Index from original reporting. Its sources are still being verified by an editor, and they are listed here once that check is complete. Until then the record carries no reviewer.