SAN FRANCISCO, July 27, 2026 (GLOBE NEWSWIRE) — VotalAI, the leadi g u time secu ity platfo m fo age tic AI, today a ou ced p oductio deployme ts of its adva ced AI Ru time Gua d ail Stack. The solutio leve ages NVIDIA NeMo Data Desig e fo la ge-scale attack data sy thesis, the NVIDIA NeMo Gua d ails toolkit, a d the NVIDIA Nemot o 3.5 Co te t Safety model stack u i g o NVIDIA AI i f ast uctu e. This e ables secu e, t usted AI deployme ts with low-late cy, o -p emises p otectio that keeps eve y toke i side the custome ’s co t olled e vi o me t.
As e te p ises a d gove me ts accele ate thei AI i itiatives, the eed fo obust u time gua d ails that p otect agai st jailb eaks, p ompt i jectio s, e codi g attacks, a d u autho ized tool use has become c itical. VotalAI’s platfo m add esses these challe ges by e fo ci g ide tity, data policies, a d tool autho izatio i eal time ac oss the AI stack, without comp omisi g model pe fo ma ce o data sove eig ty.
Key highlights of Votal’sAI
AI Ru time Gua d ail Stack with Custom Policy Ma ageme t (PDP/PEP): A policy-d ive u time with clea sepa atio betwee the Policy Decisio Poi t (PDP) a d Policy E fo ceme t Poi t (PEP), e abli g pe -te a t, pe -wo kload custom policies to be autho ed, ve sio ed, a d e fo ced i eal time without edeployi g the u time. The stack is powe ed by Votal’s custom vai35-4B-v2 AI model, a adve sa ial ha de ed gua d ail built o Qwe 3.5-4B usi g ep ese tatio e gi ee i g, su gical weight editi g, a d li ea -p obe tech iques, wo ki g alo gside NVIDIA Nemot o 3.5 Co te t Safety model a d Nemot o custom policy skills to delive 151+ detectio tech iques ac oss 19 adve sa ial attack catego ies i cludi g multi-li gual jailb eaks, e coded payloads, a d payload obfuscatio s. E d-to-e d u time late cy is app oximately 190 ms at 16k to 32k toke p ompts a d oughly 10% ove head elative to the u de lyi g LLM espo se time at co texts up to 256k toke s.NVIDIA-Powe ed T ai i g a d Post-T ai i g Pipeli e: VotalAI uses NVIDIA NeMo Data Desig e to sy thesize dive se, egio -awa e, te a t-isolated adve sa ial datasets at scale, i cludi g jailb eaks, p ompt i jectio s, e coded payloads, payload obfuscatio s, multi-tu ma ipulatio , domai -specific abuse, a d multi-li gual attack vecto s. These datasets feed Votal’s Supe vised Fi e-Tu i g (SFT) a d p efe e ce-tu i g pipeli e that p oduces the vai35-4B-v2 gua d ail model. The pipeli e combi es sy thetic DataGe , SFT o attack/defe se pai s, Rei fo ceme t Lea i g f om Huma Feedback (RLHF) a d p efe e ce tu i g d ive by huma ed-team feedback, plus co ti uous lea i g a d et ai i g cycles (app oximately eve y 3 mo ths) to p oduce te a t-specific gua d ails tu ed to each custome ’s u ique th eat su face.Be chma k a d Evaluatio Cove age: Evaluated o Ha mBe ch fo adve sa ial obust ess, co te t-safety evaluatio s ac oss 141 ha m catego ies, a d Votal’s i te al tool-misuse a d age tic RBAC evaluatio suites, with dedicated cove age fo multi-li gual attack vecto s that commo ly bypass E glish-o ly gua d ails.
Age tic Tool-Call Secu ity a d OIDC Ide tity I teg atio : Eve y age t tool i vocatio is validated i eal time agai st custom pe -te a t tool-call policies autho ed i the PDP/PEP f amewo k, with ole-based access co t ol (RBAC) fo tool autho izatio , MCP se ve validatio , data tai t t acki g, a d goal-d ift detectio . Native OIDC i teg atio with e te p ise ide tity p ovide s (Okta, Mic osoft E t a ID, Google Wo kspace) mea s eve y age t tool call ca ies the acti g use ’s scoped ide tity f om p ompt th ough executio . The stack e fo ces custome -defi ed ules o which tools ca be called, by whom, a d o which data, blocki g u autho ized tools, c oss-te a t data leakage, a d p ivilege escalatio befo e the call leaves the u time.P oductio -Ready Pe fo ma ce: SOC 2 Type II ce tified, with u time late cy as low as 10% of the u de lyi g LLM espo se time; fully suppo ts o -p em, p ivate cloud, VPC, a d ai -gapped deployme ts. Seamless i teg atio with self-hosted models, NVIDIA Nemot o ope models, Ope AI, A th opic, Amazo Bed ock, popula age t f amewo ks (La gChai , LlamaI dex, C ewAI), a d LLM p oxy/gateway laye s such as LiteLLM, Po tkey, a d Ko g, so gua d ails d op i without cha ges to applicatio code. Policies a e defi ed i simple YAML o SDKs with full CI/CD a d ve sio co t ol suppo t.Co ti uous Adaptatio via CART: VotalAI Co ti uous Adve sa ial Red Teami g (CART) platfo m uses auto omous AI easo i g age ts to co ti uously p obe gua d ails fo gaps, su face ew attack vecto s, a d automatically update policies, delive i g adaptive, p oductio -ha de ed secu ity.
“Votal AI is p oud to collabo ate with NVIDIA to b i g p oductio u time secu ity to st ategic AI i vestme ts,” said Bobby Gupta, Co-Fou de a d CEO of Votal AI. “By combi i g ou AI Ru time Gua d ail Stack, a cho ed by the custom vai35-4B-v2 model a d ep ese tatio e gi ee i g tech iques, with NVIDIA’s Nemot o 3.5 Co te t Safety model a d Nemot o custom policy skill, a d accele ated ha dwa e, we a e al eady delive i g the t ustwo thy laye that e ables safe scali g of age tic AI while p ese vi g full data sove eig ty, mi imal capability loss, a d cost-efficie t i fe e ce to ou key clie ts globally.”
VotalAI’s solutio is al eady live i p oductio e vi o me ts secu i g c itical age tic wo kloads ac oss gove me t a d e te p ises, i cludi g at leadi g data ce te ope ato s suppo ti g egulated i dust y use cases. It helps o ga izatio s meet st i ge t complia ce, eside cy, a d secu ity equi eme ts while u locki g the full pote tial of auto omous AI age ts.
VotalAI p ovides the u time secu ity platfo m fo e te p ise a d gove me t age tic AI. Its AI Ru time Gua d ail Stack a d CART platfo m delive eal-time policy e fo ceme t, ide tity co t ols (i cludi g OIDC i teg atio with Okta, Mic osoft E t a ID, a d Google Wo kspace, plus RBAC), data p otectio , a d tool autho izatio , secu i g AI i te actio s f om p ompt to tool call. Votal is SOC 2 Type II ce tified a d focused o e abli g safe, complia t AI adoptio i egulated i dust ies. Fo mo e i fo matio , visit www.votal.ai.
