Proliferation-Enabling Technologies (PETs)

• Proliferation-Enabling Technologies (PETs) are tools including AI-enhanced hardware, software, and cognitive substitutes that simplify complex nuclear bomb-design processes.
• They automate critical steps and compress tacit expertise, reducing production barriers as seen with 3D printing, digital twins, and algorithmic simulations.
• Quantitative models like the Relative Advantage Index (RAI) demonstrate PETs’ influence on detection and risk, emphasizing the need for adaptive policy and smarter DET upgrades.

Proliferation-Enabling Technologies (PETs) are defined as “tools that simplify bomb pathways: advances in hardware, software, and even cognitive substitutes for human expertise” [(Allison et al., 8 Dec 2025), p. 3845]. These technologies substantially reduce the industrial footprint, compress or substitute the tacit knowledge transfer traditionally required, or automate complex steps in the nuclear fuel cycle and warhead-design process. Their accelerating development, particularly under AI–driven paradigms, is reshaping the strategic and operational calculus of nuclear proliferation and counterproliferation.

1. Definition and Categorization of PETs

PETs comprise diverse advances across hardware, software, and AI-augmented domains. The key categories are as follows [(Allison et al., 8 Dec 2025), pp. 3842–3845]:

PET Category	Representative Examples
Hardware	Additive manufacturing (3D printing) of centrifuge parts, CNC machining for explosive lenses, affordable microcontrollers
Software	HPC-driven cascade/warhead physics simulations, digital twins for design validation, algorithmic trade-flow optimization
Cognitive-Substitution	LLMs trained on nuclear engineering literature, AI-aided reprocessing modeling, automated code generation

Hardware PETs reduce technical barriers for production and assembly, e.g., through 3D-printed maraging-steel centrifuge rotors and CNC-machined explosive lenses. Software PETs automate and scale modeling tasks, such as cascade design or masking dual-use procurement, and leverage verification-efficient digital twins. Cognitive-substitution PETs are AI-driven, including LLMs providing stepwise guidance for less-experienced actors, AI-aided fuel cycle route optimization, and automated code-generation for process control (with technical impacts cited to Frieder et al. 2023 and Almeldein et al. 2025).

2. Formalism: Relative Advantage Index (RAI)

Quantitative assessment of PETs’ impact on the balance between proliferators and detectors is formalized through the @@@@1@@@@ (RAI):

$RAI(t) = P(t) - D(t)$

• $P(t) \in \mathbb{R}^+$: Proliferator’s capability to evade detection, normalized with $P(0) = 20$.
• $D(t) \in \mathbb{R}^+$: Detection index of the international community, $D(0) = 50$.

A negative RAI corresponds to detection dominance; positive RAI infers proliferator advantage.

2.1 PET Growth: Logistic Model

PET advancement follows a logistic (sigmoid) trajectory, with growth rate $g_p$ and ceiling $P_{max}$ [(Allison et al., 8 Dec 2025), Eq. 4; Table 1]:

$$P(t) = 1 + (P_{max} - 1) \frac{\exp(-g_p t)}{1 + \exp(-g_p t)}$$

Three PET regimes are parametrized [Table 2]:

• Limited AI: $g_p=0.23$ yr$^{-1}$ ($\sim$3-year doubling)
• Disruptive AI: $g_p=0.33$ yr$^{-1}$ (25-month doubling)
• Transformative AI: $g_p=1.19$ yr$^{-1}$ (7-month doubling)

2.2 DET Improvement: Stepwise Model

Detection-Enhancing Technologies (DETs) improve discretely: under the status quo, upgrades of +5 at $t=3$ years and +4 at $t=7$ years, with $k=0.4$ logistic slope; in the “moonshot” scenario, upgrades are front-loaded (+12 at $t=1$, +10 at $t=3$, +10 at $t=6$, $k=0.6$) [(Allison et al., 8 Dec 2025), p. 3853].

2.3 Detection Probability and Cumulative Risk

Detection probability is a shifted logistic mapped from RAI:

$$Pr(D_t) = T_0 + (1-T_0)\left[1 + \exp(-k\,RAI(t) - \epsilon)\right]^{-1}$$

Parameters: $T_0=0.10$, $k=0.4$ (or $0.6$ for moonshot), $\epsilon=0$.

Hazard rate for breakout attempts is:

$$\lambda(t) = \lambda_0[1 + n \cdot \sigma(B(RAI(t)-\tau))]$$

where $\lambda_0=0.05$ yr$^{-1}$, $n=3$, $B=0.10$, $\tau=40$, and $\sigma(x) = (1+e^{-x})^{-1}$ [(Allison et al., 8 Dec 2025), Eq. 6].

Cumulative risk of undetected breakout over horizon $T$:

$R(T) = \int_0^T \lambda(t)[1 - Pr(D_t)] \, dt$

Probability of at least one undetected breakout:

$P_{breakout} = 1 - \exp[-R(T)]$

3. Scenario-Based Simulation Workflow

Allison and Herzog conduct end-to-end simulations spanning all combinations of three PET growth rates (Limited, Disruptive, Transformative AI) and two DET trajectories (status-quo, moonshot) [(Allison et al., 8 Dec 2025), p. 3852]. The methodology is as follows:

1. Set PET $g_p$ and compute $P(t)$ via logistic Eq. 4.
2. Specify DET path (baseline or moonshot), generating $D(t)$.
3. Compute $RAI(t) = P(t) - D(t)$.
4. Derive $Pr(D_t)$ from RAI using logistic mapping.
5. Calculate $\lambda(t)$, enabling or disabling opportunistic boost.
6. Numerically integrate $R(T)$ over a 10-year horizon to obtain $P_{breakout}$.

All scenarios fix $\lambda_0=0.05$ yr$^{-1}$, $T_0=0.10$, $k=0.4$ or $0.6$, $\tau=40$, $n=3$, and $B=0.10$.

4. Results: Uncertainty Band and Breakout Risk

4.1 Expansion of Uncertainty

Stepwise DET improvements contrasted with logistic PET growth generate a “band of uncertainty” that widens over time [(Allison et al., 8 Dec 2025), Fig. 2]. As PETs accelerate, this band encompasses a larger region where the proliferator’s advantage is ambiguous. Aleatory and epistemic uncertainties both contribute, increasing risk visibility only transiently after major DET jumps.

4.2 Breakout-Risk Quantification

Key cumulative 10-year breakout risks ($R(10)$, $P=1-e^{-R(10)}$, [Table 4]):

PET Regime	DET Path	$R(10)$	$P_{breakout}$
Limited AI	Status-quo	0.17	0.16
Disruptive AI	Status-quo	0.26	0.23
Transformative AI	Status-quo	0.40	0.33
Disruptive AI (+opportunism)	Status-quo	0.36	—
Transformative AI (+opportunism)	Status-quo	1.44	—
Limited AI	Moonshot	—	~$0$
Disruptive AI	Moonshot	—	~0.12
Transformative AI	Moonshot	0.39	0.32

“Moonshot” DET cuts risk by ~98% in Limited AI and ~49% in Disruptive AI scenarios; under Transformative AI, DET upgrades fail to appreciably constrain cumulative risk ($R(10) ≈ 0.39$).

4.3 Transition Thresholds

The crossover point $t^{*}$ where $P(t^*)=D(t^*)$ can be analytically solved by inverting the logistic and stepwise models. For the Limited AI, status-quo DET baseline, $t^*\approx 6.3$ yr [(Allison et al., 8 Dec 2025), p. 3853]. For high $g_p \gtrsim 1.0$ yr$^{-1}$, the PET advantage dominates soon after $t=3$ years and persists, as stepwise improvements lag.

5. Mechanisms of AI-Driven PET Acceleration

AI introduces two destabilizing mechanisms. First, logistic scaling: state-of-the-art LLMs can double their effective capability in months rather than years, shifting the PET inflection leftward in time [(Allison et al., 8 Dec 2025), Fig. 1 & 4]. Second, knowledge substitution: LLMs densely compress and automate tacit expertise, enabling personnel with less direct experience to execute critical proliferation steps with high technical fidelity [(Allison et al., 8 Dec 2025), pp. 3842–43]. This selective compression of “know-how” and “know-why” is distinct from traditional diffusion and is not easily intercepted by export controls or intelligence.

6. Governance, Verification, and Policy Interventions

The authors argue that absent rapid and sustained enhancement of DETs, or effective PET-governance, proliferation advantage will likely shift irreversibly (Allison et al., 8 Dec 2025).

6.1 Monitoring and Verification Acceleration

Policy prescriptions include a “Frontier Detection Acceleration Initiative”: multinational funding for AI-enabled analytic pipelines, satellite constellations for multispectral and thermal-infrared monitoring, rapid prototyping, and inducement prizes, coupled with public-private partnership models to minimize DET deployment latency [(Allison et al., 8 Dec 2025), p. 3858].

6.2 Adaptive Verification Regimes

Verification agencies should be restructured as agile innovation-focused organizations, leveraging modular treaty addenda and standing technical review boards. The IAEA’s Additional Protocol may require upgrades to integrate AI anomaly detection and enhanced environmental sampling, including collaborative pilots among coalitions of the willing [p. 3859].

6.3 PET Governance and Export Controls

Recommendations include extension of export controls (e.g., Wassenaar) to encompass open-weight LLM foundation models and high-resolution additive manufacturing hardware [pp. 3842, 3858], mandatory watermarking/audit logging for AI outputs, and PET-use policy development by industry, drawing parallels to biosecurity best practices.

6.4 Demand-Side Nonproliferation

Securing the nuclear fuel cycle via multilateral enrichment consortia and credible security assurances can lower the baseline hazard rate ($\lambda_0$) and opportunism multiplier ($n$). Diplomatically reducing regional tensions further decreases incentives for clandestine programs [p. 3859].

A plausible implication is that under moderate AI-driven PET growth, aggressive (“moonshot”) DET development can regain and maintain detection dominance. However, with Transformative AI rates, detection improvements saturate and only systemic PET governance and demand-reduction measures can constrain cumulative risk (Allison et al., 8 Dec 2025).