EUVIMEDEUVIMED entdecken
Crossref

IntentShield AI: A Probabilistic Self-Auditing Framework For Real-Time Risk-Constrained Alignment In Autonomous Systems

Lakshay Bhoria, Rohit Pal, Bharat Bharat

IOSR Journal of Computer Engineering · 2026-03 · 28 (3) · 91-97

DOI: 10.9790/0661-2803019197 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IOSR Journal of Computer Engineering, 2026-03. Verfasst von Lakshay Bhoria und 2 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Ensuring the safe deployment of autonomous AI systems requires mechanisms that dynamically detect and mitigate harmful outputs under uncertainty. Existing alignment tech-niques, such as Reinforcement Learning from Human Feedback (RLHF), often lack formal guarantees under adversarial per-turbations or ambiguous out-ofdistribution inputs. This paper introduces IntentShield AI, a probabilistic, multi-layered safety architecture integrating latent intent inference, Bayesian risk estimation, counterfactual simulation, and recursive selfauditing. We model AI safety as a risk-constrained stochastic optimization problem, proposing a framework that minimizes expected utility loss subject to strict safety boundaries. By formulating the environment and user interaction as a Partially Observable Markov Decision Process (POMDP), the system mathematically captures uncertainty in both user intent and environmental con-text. Theoretical analysis demonstrates bounded risk convergence by utilizing a continuous relaxation of the token logit space. We implemented the framework in PyTorch, and evaluations on 1,000 sampled prompts from standard adversarial datasets demonstrate improved robustness against modern jailbreaks (e.g., GCG attacks) and higher safety compliance compared to existing programmatic guardrails. Results averaged over three independent runs confirm high statistical significance (p < 0.01, large effect size), and ablation studies highlight the critical necessity of both the intent inference and counterfactual sandbox components.

Zitieren

APA

Bhoria, L., Pal, R., & Bharat, B. (2026). IntentShield AI: A Probabilistic Self-Auditing Framework For Real-Time Risk-Constrained Alignment In Autonomous Systems. IOSR Journal of Computer Engineering, 28(3), 91-97. https://doi.org/10.9790/0661-2803019197

Vancouver

Bhoria L, Pal R, Bharat B. IntentShield AI: A Probabilistic Self-Auditing Framework For Real-Time Risk-Constrained Alignment In Autonomous Systems. IOSR Journal of Computer Engineering. 2026;28(3):91-97. doi:10.9790/0661-2803019197

BibTeX
@article{Bhoria2026,
  author = {Bhoria, Lakshay and Pal, Rohit and Bharat, Bharat},
  title = {IntentShield AI: A Probabilistic Self-Auditing Framework For Real-Time Risk-Constrained Alignment In Autonomous Systems},
  journal = {IOSR Journal of Computer Engineering},
  year = {2026},
  volume = {28},
  number = {3},
  pages = {91--97},
  doi = {10.9790/0661-2803019197}
}
Kostenlos starten →