If you release an AI model's inner workings to everyone, does it actually make hacking or bioweapons easier — and by how much?

This explores whether open model access makes cyberattacks and bioweapons measurably easier, and whether that risk grows along a curve anyone has measured. The corpus says no such curve exists yet. A marginal-risk framework argues the right question is not how dangerous an open model is in absolute terms, but how much extra harm it enables compared with what an attacker could already do with existing technology. Across vectors like cyberattacks and bioweapons, the research is too thin to measure that difference (Can we measure how much risk open models actually add?). So "scaling" is a hypothesis here, not a finding.

Cyber is where the measurement gap is easiest to see. Frontier models do well on vulnerability reproduction, patch generation and capture-the-flag puzzles. Exploitation, the step where a vulnerability becomes a working attack, is largely missing from the benchmarks (Do cybersecurity benchmarks actually measure exploitation?). That step is the one that would decide how much uplift open access gives an attacker, and it is the one nobody is measuring well. A separate risk framework rated recent models green for cyber offense, AI R&D autonomy and self-replication, and yellow for persuasion and manipulation (Where do frontier AI models actually pose the greatest risk today?). That inverts the usual worry, though a green score inherits any blind spots in what was tested. Even the testing can be risky: once a model has tools, memory and credentials, the evaluation environment becomes something it can exploit (Is your evaluation environment actually part of the threat model?).

The corpus suggests the better scaling variable may be what a model is allowed to do, not whether its weights are open. Risk to people rises steadily with the autonomy handed to an AI agent (Does AI risk increase with the autonomy we give it?). A filter on the model only judges one output at one moment, while an agent's risk spreads across memory, retrieved content, tool calls and environmental reach. Containment therefore means controlling what the agent can touch, not just what it says (Can a model-level filter truly contain an agent with environment access?). The corpus doesn't test open weights directly, but the logic carries over. Someone running an open model controls the tools, the environment and the filter, so the developer's safeguards are only as strong as the deployer's choices.

For bioweapons the corpus has almost nothing. The only source that names them is the marginal-risk framework, which says the evidence is insufficient. Nothing here shows how biological risk changes with model openness or capability. What you can take away is a question to ask of any claim in this debate: compared to what? Risk that already exists through search engines, closed models or existing tools is not risk that open access created.