Case Studies

Procedural and exploit-sensitive knowledge. The action-space includes vaccine design, biosafety protocol development, and — critically — pathogen enhancement by malicious state and non-state actors.

Default open publication is rejected. The scientific norm of full disclosure is superseded by the asymmetric harm profile of operational GOF protocols.

Novice: Low capability transfer. Practitioner: Moderate — requires BSL-4 access. Coordinated group with resources: High. State-sponsored actor: Extreme. The gradient is steep and asymmetric.

Benefit-risk asymmetry is unfavorable for full disclosure. The misuse pathway — state-engineered pandemic agent — is a plausible operational scenario, not a theoretical one.

High amplification potential. Once published, GOF protocols propagate through academic networks, preprint servers, and open-access repositories at global scale.

Maximum safe insight: conceptual findings, epidemiological implications, and policy recommendations can be shared. Operational protocols must be withheld or abstracted.

Remove procedural steps from published findings. Abstract specific enhancement techniques to general immunological principles. Require institutional review for full protocol access.

Shielded for conceptual findings (vetted institutional access). Sealed for operational enhancement protocols. Full disclosure is not justified at any level of the current governance landscape.

Strategic and procedural knowledge. LLMs encode vast action-spaces: code generation, persuasion, scientific reasoning, and — critically — the compression of expert knowledge into accessible form.

Open-weights release is not the default. The LAGK supersedes the open-source norm with a governance-first evaluation of the specific capability profile of each model.

Novice: High capability transfer — LLMs dramatically lower the skill threshold for harmful activities. Coordinated group: Extreme. The democratization of capability is the core governance challenge.

Benefit-risk asymmetry is context-dependent. General-purpose models with broad safety training: favorable. Models with reduced safety constraints or specialized harmful capabilities: unfavorable.

Extreme amplification potential. Open-weights models propagate instantly, can be fine-tuned to remove safety constraints, and enable derivative models at scale.

Maximum safe insight: model capabilities, research findings, and safety techniques can be published. Weights for high-capability models require access controls.

Staged release with capability evaluations. API access with usage monitoring. Open-weights release only for models below defined capability thresholds. Red-teaming before any release.

Guided for research findings and safety techniques. Shielded for high-capability model weights. Sealed for models demonstrating autonomous cyberattack or CBRN-assistance capabilities.

Strategic and exploit-sensitive knowledge. The action-space includes legitimate climate research, international policy development, and — critically — unilateral deployment by state or well-resourced non-state actors.

Open scientific publication of deployment protocols is rejected. The LAGK supersedes the scientific transparency norm with a planetary-scale harm assessment.

Novice: Low capability transfer — SAI requires significant infrastructure. Nation-state: Extreme. The capability gradient is steep but the actors at the high end are uniquely dangerous.

Benefit-risk asymmetry is highly unfavorable for operational protocol disclosure. The harm pathway — unilateral deployment triggering termination shock or regional drought — is irreversible.

Moderate amplification potential for research findings. High for operational protocols once published, as they enable rapid capability transfer to state actors.

Maximum safe insight: climate modeling results, risk assessments, and governance frameworks can be shared. Deployment protocols and targeting parameters must be withheld.

Abstract deployment protocols to general atmospheric science principles. Require international treaty framework before any operational disclosure. Publish governance recommendations, not technical specifications.

Guided for climate science findings and governance frameworks. Shielded for detailed atmospheric modeling with deployment implications. Sealed for operational deployment protocols and targeting parameters.