- NVIDIA’s Open Agent Safety Platform, launched September 28, 2026, uses Sentry on BlueField-4 DPUs to quarantine misbehaving AI agents in milliseconds, outside the application layer a compromised agent can reach.
- IBM watsonx.governance monitors agents against 12 pre-loaded regulatory frameworks through evaluation nodes, and added Enforcement Tracking on August 11, 2026 to automatically capture agent evaluation metrics from watsonx Orchestrate as audit evidence.
- LangChain lacks per-agent credential concepts and an integrated kill switch, meaning teams that need identity governance in production must build those controls themselves.
NVIDIA moved AI agent security to the hardware layer on September 28, 2026, putting it on a direct collision course with software-first governance platforms like IBM’s watsonx.governance. Both tackle the same threat set, prompt injection, excessive agency, unauthorized tool execution, but from opposite ends of the stack. For teams building production agent systems, the choice comes down to where your real exposure lives.
The Agentic Security Gap
Static applications have fixed attack surfaces. AI agents don’t. They make decisions, call tools and act autonomously in dynamic environments, which means a compromised agent can do real damage before any perimeter defence fires. The OWASP 2026 guidelines reflect exactly this pressure, shifting the framing from prevention to blast radius containment: the question is no longer whether an agent will misbehave, but how much it can do before it’s stopped.
Three threat classes define the gap. Prompt injection subverts an agent’s directives through malicious input. Excessive agency occurs when an agent operates beyond its granted permissions, reaching data or systems it was never meant to touch. Unauthorized tool execution covers cases where an agent misuses legitimate access to external tools. Traditional cybersecurity tooling addresses none of these well at runtime, which is what’s driving the current wave of agent-specific security investment.
NVIDIA Goes to the Hardware Layer
The NVIDIA Open Agent Safety Platform pairs two components. OpenShell is open-source software that wraps a secure runtime boundary around agents executing on NVIDIA Vera CPUs, tracing every action and enforcing policies as they run. When a prompt injection attempt tells an agent to exfiltrate data, OpenShell catches and blocks it before it completes.
Sentry, the platform’s second component, runs as an out-of-band watchdog on NVIDIA BlueField-4 DPUs, dedicated network processors that sit outside the main compute path. Because Sentry operates independently of the application layer, a compromised agent cannot disable it. Quarantine happens in milliseconds. The platform is open-source and extends to third-party compute from Arm and Intel. Anthropic, Microsoft, Palo Alto Networks, JPMorganChase, CrowdStrike, Palantir, Salesforce and SAP are among the more than 100 organizations collaborating on the initiative.
IBM’s Lifecycle Governance Play
IBM watsonx.governance, in market since 2023, takes a software-defined approach, embedding governance into the agent workflow rather than the compute substrate. Evaluation nodes built directly into agents monitor metrics like answer relevance, context faithfulness and tool-calling accuracy, logging every agent decision for audit. The platform’s Compliance Accelerators pre-load 12 regulatory frameworks, including the EU AI Act and NIST AI RMF, mapping requirements to AI use cases automatically. The platform also carries a centralised asset catalog covering models, agents and datasets, automated onboarding workflows with policy gates before production, and integration with IBM OpenPages for GRC workflows.
In August 2026, IBM extended this with Enforcement Tracking for watsonx Orchestrate: agents in production are now evaluated on a scheduled basis, agents in development can be evaluated on demand, and metrics such as hallucination, helpfulness and toxicity are automatically captured and checked against governance thresholds, closing the gap between defined policy and proof of enforcement. For regulated industries where the governance artefact matters as much as the control itself, that automation is the main value proposition, the kind of audit trail pressure that’s already pushing firms like JPMorgan to harden their agent deployments.
What Builders Should Weigh
Governance depth is where most evaluations get shaky. Most platforms can launch an agent; fewer can prove it behaves correctly at scale. Granular access controls, per-operation audit logging and human approval gates for sensitive actions are the features that separate production-ready governance from demo-ready governance. Prompt injection detection, sensitive data leakage prevention and unauthorized tool execution blocking at runtime are non-negotiable requirements, not differentiators.
LangChain is worth calling out specifically: it offers auditable execution models that many teams value, but it currently lacks per-agent credential concepts and an integrated kill switch. Teams that need identity governance in production will need to build those controls themselves. Cost attribution, tracking token spend per agent for FinOps accountability, is becoming a fourth evaluation axis as agentic deployments scale, and neither platform has made it a first-class feature yet.
Hardware Boundary or Audit Trail
NVIDIA and IBM are not competing for the same problem. NVIDIA’s platform answers what the agent does at the compute level, it monitors and enforces at the execution layer, where hardware-level exploits and rapid containment requirements live. Its value is highest for teams running agents against sensitive systems where millisecond quarantine matters more than regulatory reporting.
IBM watsonx.governance answers how you govern what agents say and do across the whole lifecycle, across clouds and against a regulatory checklist. The two are complementary by design: NVIDIA secures the runtime; IBM documents and governs the behaviour. Which one to prioritise first usually comes down to a single question, is your compliance team being asked about the hardware boundary, or the audit trail?



