<img src="https://ml.globenewswire.com/Resource/Download/bd554f38-0905-492b-bb95-3c543dce9ddf?size=2" nborder="0" align="left" hspace="5" vspace="5" title="VotalAI logo.jpg" alt="VotalAI"/>VotalAI Runtime Guardrail Stack, featuring the vai35-4B-v2 model, now live in production securing agentic AI workloads with real-time policy enforcement from prompt to tool callSAN FRANCISCO, July 27, 2026 (GLOBE NEWSWIRE) -- VotalAI, the leading runtime security platform for agentic AI, today announced production deployments of its advanced AI Runtime Guardrail Stack. The solution leverages <a href="https://www.globenewswire.com/Tracker?data=fciwGPV0vUY90apTwPWkbFjor4GcQ9Shju0TkuwZJjb-zS4tDCRYR6ienRpAID5w_cvHIcz3XJ6Rvtk-XbLeS8eAga7MQBg4vY26NJdT9oey9WgaxvBsmcxEvWVSv2lZ-letUMr29Za4_2v7PcuawqzHDCKNq_4VSGOQpiCn0Fg=" rel="nofollow" target="_blank" title="NVIDIA NeMo Data Designer">NVIDIA NeMo Data Designer</a> for large-scale attack data synthesis, the <a href="https://www.globenewswire.com/Tracker?data=fciwGPV0vUY90apTwPWkbEnyCTUfRZyVHHRviziCjdG1Vvn3CQzfF6QKWM4HtuDX4MlEoLSoYptiOSMpNBdkCCeAiEvVJCLwR0lQeAy4QGBiq3fe1ieXg8YuH8hwmXfXazOHztfNfftHMKTOSFzf7HHDXvFUp8VcoQCDPLyCL1I=" rel="nofollow" target="_blank" title="">NVIDIA NeMo Guardrails</a> toolkit, and the <a href="https://www.globenewswire.com/Tracker?data=fciwGPV0vUY90apTwPWkbGwE0tldRtowoKXKrixMcP3qxGHTjgqKKk7Hqv1PEzAGn1i1ha0fo9OrC2hM5ZnKurIMxyB_Hp4ALxitR6VdMgn_YMxPgPCsn0WEl7AUrWJGLTxOYRq45juzees2pT5j6CKKLC0OC5ZDbgrTefyAd-8=" rel="nofollow" target="_blank" title="">NVIDIA Nemotron</a> 3.5 Content Safety model stack running on NVIDIA AI infrastructure. This enables secure, trusted AI deployments with low-latency, on-premises protection that keeps every token inside the customer’s controlled environment. As enterprises and governments accelerate their AI initiatives, the need for robust runtime guardrails that protect against jailbreaks, prompt injections, encoding attacks, and unauthorized tool use has become critical. VotalAI’s platform addresses these challenges by enforcing identity, data policies, and tool authorization in real time across the AI stack, without compromising model performance or data sovereignty. Key highlights of Votal’sAI Runtime Guardrail Stack include: <ul type="disc"><li style="margin-bottom:6pt; text-align:justify;">AI Runtime Guardrail Stack with Custom Policy Management (PDP/PEP): A policy-driven runtime with clean separation between the Policy Decision Point (PDP) and Policy Enforcement Point (PEP), enabling per-tenant, per-workload custom policies to be authored, versioned, and enforced in real time without redeploying the runtime. The stack is powered by Votal’s custom vai35-4B-v2 AI model, an adversarial hardened guardrail built on Qwen3.5-4B using representation engineering, surgical weight editing, and linear-probe techniques, working alongside <a href="https://www.globenewswire.com/Tracker?data=Bnutda98Jy6dl2Y1Rlwi9TpzSMAt8Nlk3auSG_2Y2t2Q8h_hfrqLoCvtA8cAu2stSjbSJ8O725LJpfzw6HZVJ8zIa_Ms8m8uEJN1sVKsOs78-OkTglY4ZaBmjrm3tQWULL1qhZad-R6jo5sd6d9oKk4xAnyiWbMDv4eO3qwb1G467VMDQi18fFUjaGRBQleI" rel="nofollow" target="_blank" title="">NVIDIA Nemotron 3.5 Content Safety</a> model and <a href="https://www.globenewswire.com/Tracker?data=tBxQFY2on3YkE_z20enK8zJv1G1u6FvqOXy0M_5qIrOfpsyeZlDQwrSpQOED-0bYmTZyFVfyUeJaQxrrQ_0wW975P43cGuZPstuzumdT4vnjHhL-J5w_JnopSf6GzNbHjRLKlHkU3CaDcl8K3k7hl2RkqvO7PH1MsY6rJtitLBGQwMUj0VcCv9I2z6HIQBYF" rel="nofollow" target="_blank" title="">Nemotron custom policy</a> skills to deliver 151+ detection techniques across 19 adversarial attack categories including multi-lingual jailbreaks, encoded payloads, and payload obfuscations. End-to-end runtime latency is approximately 190 ms at 16k to 32k token prompts and roughly 10% overhead relative to the underlying LLM response time at contexts up to 256k tokens.</li><li style="margin-bottom:6pt; text-align:justify;">NVIDIA-Powered Training and Post-Training Pipeline: VotalAI uses NVIDIA NeMo Data Designer to synthesize diverse, region-aware, tenant-isolated adversarial datasets at scale, including jailbreaks, prompt injections, encoded payloads, payload obfuscations, multi-turn manipulation, domain-specific abuse, and multi-lingual attack vectors. These datasets feed Votal’s Supervised Fine-Tuning (SFT) and preference-tuning pipeline that produces the vai35-4B-v2 guardrail model. The pipeline combines synthetic DataGen, SFT on attack/defense pairs, Reinforcement Learning from Human Feedback (RLHF) and preference tuning driven by human red-team feedback, plus continuous learning and retraining cycles (approximately every 3 months) to produce tenant-specific guardrails tuned to each customer’s unique threat surface.</li><li style="margin-bottom:6pt; text-align:justify;">Benchmark and Evaluation Coverage: Evaluated on HarmBench for adversarial robustness, content-safety evaluations across 141 harm categories, and Votal’s internal tool-misuse and agentic RBAC evaluation suites, with dedicated coverage for multi-lingual attack vectors that commonly bypass English-only guardrails.</li></ul> <ul type="disc"><li style="margin-bottom:6pt; text-align:justify;">Agentic Tool-Call Security and OIDC Identity Integration: Every agent tool invocation is validated in real time against custom per-tenant tool-call policies authored in the PDP/PEP framework, with role-based access control (RBAC) for tool authorization, MCP server validation, data taint tracking, and goal-drift detection. Native OIDC integration with enterprise identity providers (Okta, Microsoft Entra ID, Google Workspace) means every agent tool call carries the acting user’s scoped identity from prompt through execution. The stack enforces customer-defined rules on which tools can be called, by whom, and on which data, blocking unauthorized tools, cross-tenant data leakage, and privilege escalation before the call leaves the runtime.</li><li style="margin-bottom:6pt; text-align:justify;">Production-Ready Performance: SOC 2 Type II certified, with runtime latency as low as 10% of the underlying LLM response time; fully supports on-prem, private cloud, VPC, and air-gapped deployments. Seamless integration with self-hosted models, NVIDIA Nemotron open models, OpenAI, Anthropic, Amazon Bedrock, popular agent frameworks (LangChain, LlamaIndex, CrewAI), and LLM proxy/gateway layers such as LiteLLM, Portkey, and Kong, so guardrails drop in without changes to application code. Policies are defined in simple YAML or SDKs with full CI/CD and version control support.</li><li style="margin-bottom:6pt; text-align:justify;">Continuous Adaptation via CART: VotalAI Continuous Adversarial Red Teaming (CART) platform uses autonomous AI reasoning agents to continuously probe guardrails for gaps, surface new attack vectors, and automatically update policies, delivering adaptive, production-hardened security. </li></ul> “Votal AI is proud to collaborate with NVIDIA to bring production runtime security to strategic AI investments,” said Bobby Gupta, Co-Founder and CEO of Votal AI. “By combining our AI Runtime Guardrail Stack, anchored by the custom vai35-4B-v2 model and representation engineering techniques, with NVIDIA’s Nemotron 3.5 Content Safety model and Nemotron custom policy skill, and accelerated hardware, we are already delivering the trustworthy layer that enables safe scaling of agentic AI while preserving full data sovereignty, minimal capability loss, and cost-efficient inference to our key clients globally.” VotalAI’s solution is already live in production environments securing critical agentic workloads across government and enterprises, including at leading data center operators supporting regulated industry use cases. It helps organizations meet stringent compliance, residency, and security requirements while unlocking the full potential of autonomous AI agents. About VotalAI, Inc VotalAI provides the runtime security platform for enterprise and government agentic AI. Its AI Runtime Guardrail Stack and CART platform deliver real-time policy enforcement, identity controls (including OIDC integration with Okta, Microsoft Entra ID, and Google Workspace, plus RBAC), data protection, and tool authorization, securing AI interactions from prompt to tool call. Votal is SOC 2 Type II certified and focused on enabling safe, compliant AI adoption in regulated industries. For more information, visit www.votal.ai. Contact: <a href="https://www.globenewswire.com/Tracker?data=h7Pq2RGj_UHo_uBiME2Zam0YSUaB0pX5v4lg6vtmxA6gl8Rx2oeWWHxNjgYsayM9Mkyom6B8i-GwQ5Z95QhiYg==" rel="nofollow" target="_blank" title="info@votal.ai">info@votal.ai</a> <img alt="" src='https://ml.globenewswire.com/media/ZTk3YjU0OGQtODc4Mi00NjVhLTk1ZDYtNWNhMGEwZmUwM2NjLTEzMjE4NDItMjAyNi0wNy0yNy1lbg==/tiny/Votal-AI-Inc.png' referrerpolicy='no-referrer-when-downgrade' />

GlobeNewswire Inc.

VotalAI Delivers Production-Grade Runtime Security for Strategic AI Investments Powered by NVIDIA Nemotron

Latest news from the world of business and finance – both local and global

MONEY FM 89.3 is Singapore's first and only Business and Personal Finance station, featuring local and international business and financial news.

MONEY FM 89.3

Discover newsletters from our DJs on LinkedIn