Anthropic Limits Mythos Access to Mitigate Cyber Risk

Anthropic has limited initial access to its new Mythos model to a small group of corporate partners and is running a defenders-first testing program. Canada's AI minister Evan Solomon praised the decision, saying, "Working with defenders first, rather than releasing this new model broadly, is the responsible path and gives people protecting critical systems a head start." The firm has warned Mythos may be powerful enough to enable cyberattacks if not stress-tested against real defenses.

Anthropic is gating Mythos behind a controlled access program dubbed "Project Glasswing." Initial participants include:

• •JPMorgan Chase
• •Amazon
• •Apple

Participants will test the model against their own systems to surface misuse vectors and attack patterns prior to any wide release. U.S. authorities, including the Treasury Department technology team, are seeking access to analyze vulnerabilities, and the Federal Reserve and Canadian financial resilience groups are already discussing implications. Anthropic's public posture frames the activity as adversarial testing and hardening, not open distribution.

This is an exemplar of staged, risk-aware model rollout for high-capability systems. The defenders-first model aligns developer incentives with critical-infrastructure operators and regulators, reducing the window in which bad actors can weaponize a model before defenders can adapt. For practitioners, this raises expectations for: coordinated disclosure protocols, joint red-team exercises, and pre-release security engagements between model vendors and sector regulators.

Whether other frontier-model developers adopt similar gating and how standardized threat-sharing and testing frameworks emerge across finance, telecom, and energy. Also watch whether regulators demand formal access or certification processes for high-risk models.