NVIDIA launches Sentry, a hardware layer that watches and shuts down rogue AI agents
In short: NVIDIA has detailed Sentry, a new security framework that sits outside an AI agent's reach, enforcing policy and watching behavior from a dedicated hardware trust domain on BlueField-4 DPUs. The idea is simple: even if an agent is compromised or goes off-script while using tools and credentials, Sentry keeps watching and can cut it off because it doesn't depend on the host OS the agent runs on. Two components do the work — DOCA Tower, a proxy that inspects and authorizes every agent network request, and DOCA Argus, which reads live memory to understand what's actually running. When something looks wrong, the system can quarantine the agent within milliseconds, without needing the agent or the host to cooperate.
This summary was generated automatically by AI from NVIDIA Developer's publication. It is our own text, not a copy of the original — facts, figures and quotes belong to the source, linked above and below.
What changed?
- 1Runs on a separate hardware trust domain (BlueField-4 DPU), fully isolated from the host OS the agent executes on
- 2DOCA Tower decrypts and inspects agent traffic, checking requests against Open Policy Agent (Rego) rules in under a millisecond
- 3Access policies can require matching container hashes, verified software-bill-of-materials entries, and approved execution paths before allowing a connection
- 4DOCA Argus performs agentless live memory analysis of containers, processes, and resource usage to flag suspicious activity such as data exfiltration or unauthorized file deletion
- 5Dedicated AI classifier models scan agent requests and responses, either immediately blocking harmful traffic or flagging it for review after the fact
- 6Covers agent-to-agent and MCP traffic specifically, including detection and redaction of leaked API keys, tokens, and passwords
- 7Supports SPIFFE/SPIRE workload identity and keeps tamper-evident, attested logs for incident investigation even if an agent escapes its sandbox
Why it matters
As companies give AI agents real tools, credentials, and API access, a single agent that's tricked or hijacked can act outside its intended boundaries while still appearing to complete its task — standard software sandboxing alone may not catch this. A hardware-level watchdog that can't be disabled by the agent itself is a meaningful step for securing autonomous agent deployments, particularly ones using MCP for tool calling.
What it means for AI agents and contact centers
This is infrastructure-level security tied to NVIDIA's BlueField-4 DPU hardware, so it isn't something to deploy directly, but it signals where the industry is heading: treating agent-to-tool and MCP traffic as something that needs independent inspection, policy enforcement, and audit trails. Worth keeping in mind when designing logging, access control, and anomaly detection around any agents that call external tools or APIs on behalf of callers.
Sources
- NVIDIA DeveloperOfficialPrimary sourceOriginal article →„Inside NVIDIA Sentry: Secure AI Agents at Runtime“21 Oct 2026, 12:52
- Published by source
- 21 Oct 2026, 12:52
- Found by our system
- 21 Oct 2026, 13:17
- Summary generated
- 21 Oct 2026, 13:19
This article was written by AI from the original source. Facts, numbers and prices come from the source; missing values are marked “Not specified”. Legal notice, copyright and privacy