Anthropic and OpenAI are taking divergent approaches to deploying advanced AI systems with sensitive security applications. Anthropic chose a restricted path, limiting Claude Mythos access to a small corporate group through Project Glasswing, treating frontier security capabilities as a controlled asset. OpenAI released GPT-5.5 to the public, making comparable hacking and security tools broadly available. Critics call this release "Mythos-like hacking, open to all," signaling concerns that democratizing frontier AI security capabilities bypasses the safety guardrails Anthropic imposed.
The tension reflects a fundamental disagreement about risk management in AI deployment. Restricting access preserves control and allows companies to monitor how powerful models are used, particularly for applications that could enable cyberattacks or exploit vulnerabilities at scale. Public release accelerates innovation and availability but removes oversight mechanisms. When security-focused capabilities become widely available, malicious actors gain equal access to beneficial users.
The AI Security Institute is tracking this tension as a key trend. Their involvement suggests the government views this divergence as potentially problematic for national security and critical infrastructure protection. If advanced hacking tools become commoditized, defenders struggle to stay ahead. If tools remain restricted, some argue it slows legitimate security research and creates moats for well-funded companies.
This May 2026 snapshot captures an industry inflection point. OpenAI's decision to go public suggests confidence in GPT-5.5's safety or pragmatism about competitive advantage. Anthropic's selective release signals either greater caution or a bet that scarcity and control create defensible business value. Neither approach is obviously correct. History shows that restricted technologies often leak or spawn competitors who release openly. Public release can fuel dual-use harms faster than the community can defend.
The real issue: security capabilities in advanced AI are not neutral tools. How they're deployed