Safety

Dual-use

ActivePublished: 6 August 2026Updated: 6 August 2026Published

Key innovation

A framing of AI in which the same capability or model can serve both beneficial (civilian, research, defensive) and harmful ends - requiring risk management rather than a simple ban.

How it works

1) A capability/model with dual-use potential (beneficial and harmful) is identified. 2) Its risk is assessed: capability evaluations (evals), red teaming, analysis of misuse scenarios and thresholds. 3) Measures proportionate to the risk are chosen: guardrails, filtering, access controls, tiered access, an open-vs-closed weights decision. 4) Governance is applied: responsible disclosure, usage policies, monitoring and regulatory compliance.

Problem solved

It helps reason about AI technologies that cannot be simply classified as 'good' or 'bad'. Instead of a ban, the dual-use framing steers toward risk management (evals, guardrails, access controls, governance).

Components

Dual-use capabilityObject of assessment

The same function with beneficial and harmful value (e.g. vulnerability discovery, bio/chem knowledge).

Risk and capability assessmentRisk diagnosis

Measuring dangerous capabilities and misuse scenarios (evals, red teaming, thresholds).

Official

MitigationsRisk reduction

Guardrails, filtering, access control, tiered access, open-vs-closed weights decision.

Official

Governance and policyManagement framework

Responsible disclosure, usage policies, monitoring, regulatory compliance.

Official

Implementation

Implementation pitfalls

Over- or under-restrictionHigh

Too-strict measures destroy beneficial uses; too-weak ones fail to prevent misuse.

Fix:Eval-based proportionality, tiered access, iterative adjustment.

Irreversibility of open weightsHigh

Once weights are released, access cannot be revoked and server-side guardrails cannot be added.

Fix:Careful pre-release assessment, staged/API-first access where appropriate.

Research publication dilemmaMedium

Publishing effective tools/attacks can facilitate misuse (info-hazard).

Fix:Responsible disclosure, redacting details, coordinating with providers.

Evolution

2018

Malicious-use-of-AI report

The report 'The Malicious Use of Artificial Intelligence' (2018) formalizes concern about the dual-use of AI.

2023

Dual-use of frontier models

Inflection point

With LLMs/GenAI, dual-use becomes central to releases, red teaming and regulation (e.g. the US AI executive order, capability evaluations).

2025

Capability thresholds and release policies

Formalization of dangerous-capability thresholds and deployment policies (responsible scaling, tiered access) in labs and regulation.

Hyperparameters (configurable axes)

Risk domainHigh

Cyber, bio/chem, disinformation, fraud, autonomy, etc.

Access modelCritical

Open vs closed weights, tiered access, API vs download.

Mitigation strengthHigh

Proportionality of guardrails/controls to the assessed risk.

Computational complexity

Time complexity: Nie dotyczy (koncept polityki/bezpieczenstwa, nie algorytm). Space complexity: Nie dotyczy.

Compute bottleneck

Credible risk assessment without losing benefits

The challenge is not compute but credibly measuring dangerous capabilities and choosing proportionate measures - without blocking useful applications (the safety/usefulness balance).

Depends on

Jakosc ewaluacji zdolnosciProporcjonalnosc srodkow