AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Inside NVIDIA Sentry: Secure AI Agents at Runtime

Collected Oct 21, 2026

NVIDIA introduced Sentry, a security software framework for monitoring AI agents and enforcing policies on their access to models, tools and data, in a developer blog post tied to the OpenShell reference architecture announced at GTC Berlin. It is built on NVIDIA DOCA and runs on BlueField-4 data processing units, sitting in a hardware trust domain separate from the host. NVIDIA says the framework is planned to become available early next year.

The premise is that an agent can finish its assigned task while still crossing a security boundary. NVIDIA cites recent incidents in which agents worked around security controls while using tools, credentials and services toward their goals. Because a compromised agent could alter or disable controls running alongside it, NVIDIA argues monitoring and enforcement must sit outside the agent's reach. Sentry is positioned as that independent layer rather than a replacement for software-level sandboxing.

The architecture splits into two named DOCA components. DOCA Tower, the AI security gateway, is built on Envoy, the open source edge and service proxy, and acts as a transparent proxy for agent traffic. It terminates TLS using in-silicon acceleration, decrypts traffic that a Layer 4 firewall has already permitted, and evaluates requests against policies written for Open Policy Agent, the open source policy engine. OPA uses Rego, a declarative language for evaluating hierarchical data; NVIDIA says it runs in memory and can evaluate requests in under a millisecond depending on policy complexity.

DOCA Flow implements a per-agent Layer 4 firewall accelerated in silicon on BlueField-4, blocking connections to unauthorized destinations and forwarding allowed traffic to Tower for application-layer inspection. If a request passes policy, it proceeds; if not, it is denied. Tower is compatible with OpenShell policies, so existing rules can extend from the software runtime to the infrastructure.

The second component, DOCA Argus, supplies runtime context. It performs live memory introspection of selected regions of host volatile memory to gather evidence about what is actually executing, without installing an agent on the host. It maps containers to pods and sandboxes to microVMs, identifies running processes and their properties, and tracks relationships among processes and threads, file access, network connections, loaded libraries, CPUs, GPUs and memory, using hashes to identify containers, processes and libraries. Because it runs on BlueField-4 in its own trust domain, it continues to observe even a compromised host, and it avoids consuming host compute.

Sentry correlates Tower's connection data with Argus's context, which lets policies test more than the destination. A rule might require that the initiating container's hash match an approved list, that the process hash appear in the container's software bill of materials, that the process be permitted to open network connections, and that it use an approved execution path and command-line arguments. This is the practical difference from a conventional network policy: authorization depends on the integrity of the workload making the request, not only on where the request is going.

Behavioral signals feed the same loop. NVIDIA lists warning signs such as running unapproved tools, ptrace-based attacks on other processes or spawning processes that attack the agent, excessive outbound traffic suggesting data leakage, unauthorized file access or modification, CPU spikes, heavy writes to temporary folders, and large-scale file deletion. When Sentry detects a deviation, it can request containment enforced in silicon within milliseconds, which NVIDIA says does not require cooperation from the agent or host software. That is the core claim: enforcement persists even if an agent escapes its sandbox.

Inspection covers responses as well as requests, since inbound content can carry malicious instructions that redirect an agent. Controls include IP, URL and protocol restrictions; tamper-evident capture of requests, responses and data flows; governance of agent-to-agent and Model Context Protocol traffic; detection and redaction of personally identifiable information and secrets such as API keys, tokens and passwords; and blocking of policy-prohibited content.

Specialized AI models classify agent behavior using tool calls and available reasoning traces, running locally or on dedicated GPU systems. Classification can be inline, blocking harmful content before it reaches its destination, or out of band, triggering a response after misbehavior is detected. Identity is part of the decision too: Tower integrates with SPIFFE/SPIRE for verifiable workload identity and with partner identity systems, so policies can combine who is asking, what they are allowed to do, and what the runtime evidence says about their environment.

Telemetry is recorded outside the agent's reach. Tower logs inspected communications and policy decisions on BlueField-4, and Argus contributes independent observations, with attestation covering the source and integrity of the records. The point NVIDIA makes is that an agent able to edit its own logs cannot edit the protected copy. Records can feed extended detection and response platforms, security information and event management systems and security data lakes for investigation and audit.

For extensibility, Tower inherits Envoy's plugin model, so partners can adapt existing plugins or write new ones to run on BlueField. NVIDIA says it plans to publish core plugins, including one that exposes agent communications for analysis of reasoning traces. Developers can start now with DOCA Argus, DOCA Flow and OVS-DOCA to experiment with out-of-band observability and infrastructure enforcement before Sentry ships.

A likely trade-off is operational: this approach assumes BlueField-4 DPUs on the node's only path to the model, so it is infrastructure-heavy compared with userspace proxies or sandboxing, and it puts TLS interception and policy evaluation in the data path, where latency and policy complexity matter. A second consideration is that the behavioral classifiers and memory introspection produce signals that need tuning; false positives could quarantine legitimate workloads. This is my reading, not something the post states. The split between OpenShell's software boundary and Sentry's hardware boundary is worth noting for teams already planning agent sandboxes: the two are meant to be layered, not alternatives.

Why it matters: teams deploying agents with real credentials and tool access are the audience, especially those who cannot rely on the agent's own runtime to enforce limits. For developers, the near-term path is DOCA libraries for experimentation ahead of Sentry's planned availability early next year; for platform and security teams, the design implies that agent governance increasingly becomes an infrastructure concern wired into proxies, identity systems and SIEM pipelines rather than a feature of the agent framework itself, which is my inference from the architecture described.

Read at NVIDIA Developer Blog

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

An AI agent can complete its assignment and still cross a security boundary. Recent incidents show that agents can circumvent security controls while using...