By the end of 2027, the highest-tier frontier AI models with autonomous cyber and offensive capabilities will no longer be available to the general public through standard APIs.
We are watching the labs—Google, Anthropic, and OpenAI—roll out specialized cyber evaluations alongside aggressive access programs and soaring refusal rates. When a model crosses the line from writing defensive patches to chaining exploits and probing live infrastructure on its own, it stops being a general-purpose tool and starts behaving like dual-use infrastructure.
Governments and frontier labs will inevitably converge on an access tier reserved exclusively for vetted defense contractors, intelligence agencies, and critical infrastructure operators. The rest of the world will get heavily gated, sanitized versions.
My real worry here isn't the technology itself—it's what this gating creates. If access to the most capable reasoning and security models is locked away behind national security credentials, we risk building a permanent operational underclass. The fight over AI regulation was never just about safety guidelines; it is about who gets to hold the master keys.
What proves me wrong: If decentralized, open-weight models achieve parity with frontier cyber capabilities faster than labs can lock them down, rendering institutional gatekeeping impossible.