Skip to content

 
 
// Product: enforce Runtime Controls

Autonomy Without Oversight = Exposure

Agents don't wait for a human to approve every step — and neither should your security. Prediction Guard enforces runtime controls at the moment of action: every model call, every tool invocation, every piece of data an agent touches is checked against your rules in real time, not reviewed after the fact. Deny-by-default. No standing exceptions. No action goes through unchecked.

Real-time policy enforcement on agent actions in the Prediction Guard console
// Aligned to
NIST AI RMF OWASP ISO 42001 HIPAA SOC 2EU AI ActITARAI Action Plan
// Input / Output Safeguards

Deny-by-Default Agent Behavior Controls, on Every Action

Every model interaction, tool call, and data access request is evaluated against your runtime controls before it executes (not sampled, not logged for later review). If an action isn't explicitly permitted, it doesn't happen.

Why it matters: In the July 2026 Hugging Face intrusion, the compromised agent never had to break a rule. It exploited the absence of one. It stayed narrowly focused on its assigned task while violating every unstated boundary around it. Deny-by-default closes that gap: an agent can't wander into unauthorized territory that was simply never denied.

Deny-by-default Agent Behavior Controls configuration
Enforcing industry compliance frameworks at runtime
// Industry Best Practices

Custom Controls, or Industry Frameworks Out of the Box

Define your own runtime controls, or enforce established frameworks like NIST 600-1, OWASP recommendations, and the EU AI Act directly at runtime. Controls apply uniformly across every model, every provider, and every agent in your AI system.

Why it matters: Compliance shouldn't be a quarterly audit that discovers what already went wrong. When frameworks are enforced live, violations are prevented, not just documented after the damage is done.

// MCP Control

Least Agency, Enforced at the Tool Level

Beyond least privilege for users and systems, Prediction Guard restricts the scope of what each individual MCP server can do (a database tool gets read-only queries, an email tool gets no send rights unless explicitly scoped, etc.). Each agent acts only on the authority delegated to it, never the full reach of the account it runs under.

Why it matters: Traditional access controls can't stop an agent from misusing tools it's technically authorized to use. Least agency constrains the action itself, not just the account behind it.

Screenshot from 2026-05-13 13-36-39
Immutable audit log of AI security events
// Curate your Supply Chain

Immutable Audit Logs for Every Agent Action

Every control decision, blocked action, and violation is written to append-only, tamper-resistant storage. No single point of failure (or single compromised credential) can alter the record. Every entry ties back to a specific agent identity, timestamp, and triggering request.

Why it matters: During the Hugging Face incident, defenders relied on reconstructing over 17,000 recorded events to separate real attacker activity from decoys and establish what actually happened. That reconstruction is only trustworthy if the log itself couldn't have been altered by the same actor who compromised the environment. An immutable trail means your forensic record survives the incident it's documenting (and holds up for compliance audits, regulatory review, and legal proceedings, not just internal investigation).

Evidence bundles: for any incident, the log becomes a single evidence bundle: what the agent attempted, which control fired, what was contained, and the chain back to the accountable human. It is the record an auditor, regulator, or underwriter needs after the fact.

// Two layers of runtime control

Runtime Controls That Are Always On, Not Bolted On

A filtered agent isn't a controlled one. Prediction Guard enforces two layers of runtime control on every agent action: what goes into and comes out of each model, and what the agent is allowed to do with the tools and systems it can reach.

Layer 1

Component Input/Output Controls

Inspect every prompt, response, and tool payload for PII, prompt injection, and unsafe output before it moves on.

// PII

PII Processing

Prevent unauthorized disclosure of PII. Block or log when PII flows into or out of models or MCP servers. Remediate these instances by masking, faking, or replacing PII within user inputs or AI outputs.

// INJECTION

Prompt Injection Detection

Detect and block jailbreaking, direct prompt injection, and indirect prompt injection attacks across every model and agent in the system.

// OUTPUTS

Handle Unsafe AI Outputs

Enforce the detection and logging of unsafe model outputs, including toxicity, lack of grounding, malicious URLs, and system prompt leakage.

Layer 2

Agent Behavior Controls

Decide which tools an agent can call, in what sequence, and against which systems, under authority delegated from an accountable human.

// DENY-BY-DEFAULT

Deny-by-Default Actions

Every tool call and data access request is checked before it executes. If an action isn't explicitly permitted, it doesn't happen.

// LEAST AGENCY

Delegated Authority, Scoped per Tool

Scope what each MCP server and tool can do, so an agent acts only within the authority delegated to it, not everything its service account can reach.

// BASELINES

Framework Baselines

Quickly align your runtime controls to frameworks like NIST AI RMF, NIST 600-1, OWASP LLM Top 10, and OMB M-26-04.

Across both layers

Observability and Evidence

Every control decision is recorded, so you can prove what was permitted, what was blocked, and why.

// AUDIT

Immutable Audit Logs

Maintain a tamper-proof log of every control decision with attribution information, timestamps, and input/output data for compliance.

// EVIDENCE

Evidence Bundles

Package what an agent attempted, which control fired, and what was contained into a record an auditor or regulator can use.

// SIEM

SIEM / SOAR Integration

Stream structured AI events to Splunk, Grafana, Datadog, or any other SIEM/SOAR system used within your organization.

// Questions

Frequently Asked Questions

01 What is an AI governance platform? +

An AI governance platform is software that enforces policy, tracks risk, and produces audit-ready evidence across an organization's AI systems, models, and agents. Most platforms route AI traffic to external infrastructure to apply policy. The enforcement decision and the resulting audit log both leave your environment.

Prediction Guard runs as a self-hosted control plane instead. Runtime controls are enforced before a model call completes, and the resulting audit logs are generated inside your own infrastructure and consumed by your SIEM.

02 What should I look for in AI governance tools? +
03 How is AI governance software different from AI security software? +
04 Do AI governance solutions require sending data to a third party? +
 
// Ready to Talk?

Runtime Controls That Are Always On.

Stop treating compliance as a quarterly audit. Prediction Guard enforces runtime controls on every model, tool, and agent action.