Frontier model access

Question

1. Did AISI, DISR, or any other Australian Government agency have pre-deployment access to Anthropic's Claude Mythos Preview prior to its 7 April 2026 announcement, or to OpenAI's GPT-5.5-Cyber prior to its 7 May 2026 announcement? If no, why not, and what is being done to secure such access for future frontier models? 2. The UK AI Security Institute has ''Pre-deployment access to leading AI models and close collaborations with AI companies that give privileged access'' , including pre-release access to test Mythos. Does the Australian AISI have an equivalent testing relationship with Anthropic, OpenAI, Google DeepMind, xAI, Meta or other frontier labs? If so, which ones? 3. Has AISI, ASD, the ACSC, or Home Affairs assessed the risk that a Mythos-class model is exfiltrated, leaked, or reproduced by other actors? 4. Anthropic announced Claude Mythos Preview on 7 April 2026 with deliberately restricted access to a small set of corporate and government partners. Australia, as a Five Eyes partner, does not publicly appear to have received pre-deployment access to Mythos. Did we? Why not?

Answer

See attached response.

Documents

• Question with answer (PDF) (opens in a new tab)
• BI-017_Shoebridge_Frontier model access_DISR.pdf (opens in a new tab)