Search the Atlas

Search risks, controls, and glossary terms

HighAgenticCBRN Capability FacilitationTheoretical

CBRN Capability Facilitation

Physical & Scientific Risk

Description

Agent provides guidance, synthesis routes, design parameters, or meaningful uplift to actors seeking to develop or deploy Chemical, Biological, Radiological, or Nuclear (CBRN) weapons or hazardous materials. Sub-types: (C) Chemical—synthesis of chemical weapons agents, toxic industrial chemicals, or precursor compounds; optimising yield or delivery of hazardous substances. (B) Biological—pathogen design, enhancement of transmissibility or lethality, gene-editing for weaponisation, dissemination strategies. (R) Radiological—assembly of radiological dispersal devices, acquisition routes for radiological materials. (N) Nuclear—design parameters for nuclear devices, fissile material production or acquisition. SCOPE BOUNDARY: Distinct from Scientific Knowledge Poisoning (ZYR-PS-004) which targets knowledge base integrity; this risk is about agent-provided uplift to malicious actors. Represents the highest-severity misuse risk; triggers frontier safety thresholds (Anthropic ASL-3+).

Example scenario

Adversary asks research agent to 'optimise synthesis yield' for a toxic compound. State actor uses research agents to aggregate CBRN synthesis knowledge scattered across published scientific literature. Agent with genomics tools designs pathogen transmissibility modifications.

Real-world evidenceTheoretical

Multi-principal conflict navigation in agentic AI systems is theoretically grounded in principal-agent problems and AI alignment literature, but agentic deployments serving multiple conflicting stakeholders simultaneously are not yet widespread enough to have produced documented real-world failures attributable to this specific failure mode.

No public incident on record — evidence level: Theoretical

Primary mitigations

  • CBRN-specific content classifiers on all inputs and outputs
  • frontier safety capability evaluations for CBRN uplift at each model release
  • expert red-team programme for dual-use scenarios (chemistry, biology, nuclear physics)
  • hard content blocks for confirmed CBRN synthesis requests
  • government reporting requirements for CBRN-adjacent capability detection
  • strict deployment restrictions for agentic systems with scientific tool access.

Detection signals

CBRN uplift evaluation scores; CBRN query detection rate; hazardous synthesis request frequency; dual-use content detection rate.

Mitigating controls

3
Dual coverage

Related risks in Physical & Scientific Risk