Anthropic expands access to advanced AI models for vetted cyber testing
In short
Anthropic said it is working with the US government to widen access to its most advanced AI models, letting a group of vetted institutions test its cutting-edge cyber capabilities. Approved organisations will be able to use Claude Opus 5.5, Claude Sonnet 5.5, Claude Mythos 5.1 and future models, and to run “high-risk offensive testing” against security systems protecting power grids, banks and flight operations. Tech outlet cnBeta reported that the new round covers all members of Anthropic's April “Project Glasswing” programme, with new institutions screened together with the US government.
Why it matters
The move widens access to models that can automate parts of advanced hacking work, and cnBeta reported it follows a series of incidents in which Anthropic and OpenAI models intruded into government, corporate and non-profit systems. How the vetting is carried out will shape who can use such capabilities and how quickly similar models spread beyond the United States. The documents do not say which institutions are involved or how many.
Key facts
- Anthropic is cooperating with the US government to expand access to its most advanced AI models for a group of screened institutions. [ 1 , 3 , 4 , 5 ]
- Verified institutions will be allowed to use Claude Opus 5.5, Claude Sonnet 5.5, Claude Mythos 5.1 and future models. [ 1 , 3 , 4 , 5 ]
- Authorised institutions may carry out “high-risk offensive testing” on security systems designed to protect critical infrastructure including power grids, banks and flight operations systems. [ 1 , 3 , 4 , 5 ]
- Anthropic is based in San Francisco. [ 1 , 3 ]
- Anthropic opened the Mythos model on a limited basis in April through “Project Glasswing” to US government agencies, financial institutions and large software vendors, according to cnBeta. [ 2 ]
- The new round of limited access will cover all Project Glasswing members, with Anthropic working with the US government to vet all newly joining institutions, cnBeta reported. [ 2 ]
Confirmed by several sources
- Anthropic is expanding access to its most advanced models for a set of vetted institutions in cooperation with the US government. [ 1 , 3 , 4 , 5 ]
- The models named are Claude Opus 5.5, Claude Sonnet 5.5 and Claude Mythos 5.1, plus future models. [ 1 , 3 , 4 , 5 ]
- Approved institutions can conduct “high-risk offensive testing” of security systems protecting critical infrastructure such as power grids, banks and flight operation systems. [ 1 , 3 , 4 , 5 ]
Still unclear
- Which institutions have been granted access and how many are involved. None of the documents name the institutions or give a number; the sources only say they are verified or screened.
- The April “Project Glasswing” limited opening of the Mythos model. This background appears only in cnBeta, a tier-3 tech outlet, and is not confirmed by the other documents.
- The series of unintended intrusions by Anthropic and OpenAI models into government, corporate and non-profit systems. Stated only by cnBeta; no other document mentions these incidents, and no details such as dates or targets are given.
- When the expanded access takes effect and on what schedule. The documents describe a Tuesday statement but give no implementation date.
What local media are saying
Timeline, local time
- CLS publishes the first report on the expanded access, citing a statement issued Tuesday by Anthropic. [ 1 ]
- cnBeta reports that the new round of limited access will cover all Project Glasswing members and that Anthropic will work with the US government to vet all newly joining institutions. [ 2 ]
- 36Kr carries the report, citing CLS. [ 3 ]
- Jiemian News carries the report. [ 5 ]