IBM’s Praxis Policy Engine: What AI Agent Security Controls Can Actually Do
IBM and Red Hat describe an open-source AI gateway policy engine. The architecture is concrete, but bypass resistance and real-world assurance remain research questions.
Original conceptual illustration generated for DeepTech; not a photograph or image of the reported experiment.
An AI agent connected to company systems introduces a security problem that a chatbot alone does not. Once software can invoke tools, retrieve restricted files and act on behalf of users, organizations need enforceable answers to who may do what, with which data and under whose authority. IBM Research’s 9 November 2026 technical note on the Praxis Policy Engine (PPE) describes one open-source response: moving authorization, identity checks and audit decisions into shared infrastructure around the agent.
The distinction between an available security mechanism and a proven secure deployment matters. A gateway can enforce a configured rule; that does not establish that every unsafe action is blocked in every environment.
What IBM actually announced
In IBM Research’s technical note, IBM says it contributed the Praxis Policy Engine in collaboration with Red Hat. The open-source Praxis proxy can apply policy decisions to conventional service calls and AI traffic. IBM also describes the project’s involvement with the Cloud Native Computing Foundation Sandbox, which represents a community project milestone, not certification of a product’s security.
The separate Praxis Policy Engine repository documents identity resolution, authorization, scoped delegation, field redaction, approval workflows and audit records. This source provides inspectable implementation material; it is still not an independent penetration test of an enterprise installation.
Identity and delegation are the first boundaries
Consider an agent instructed to summarize financial data and send the result to a customer. The legitimate user, the agent runtime, each tool and the destination service may have different permissions. An agent should not inherit unrestricted access just because it was launched by an administrator. Praxis describes identity checks covering users, agents and workloads, and token exchanges that can scope delegated credentials to specific upstream services.
For buyers, the critical questions are operational: where is the identity verified, what happens when it expires, how are delegated credentials revoked and can a tool be called with privileges broader than those granted for the request? The mere presence of an identity feature does not answer those questions without deployment evidence.
Policy checks at tool and data boundaries
IBM positions gateway enforcement as a complement to host isolation. Sandboxing may limit a process’s local files and system calls, whereas a gateway can decide which network service or tool call is permitted. Such a separation is useful because an agent may encounter malicious or misleading content through retrieved material, even when the model itself has received a legitimate high-level instruction.
Policies also need to consider the direction of data movement. A tool response can contain confidential fields; a later tool may send them elsewhere. The repository describes redaction and information-flow controls, but these must be configured and tested with representative payloads. Enterprise teams should record both correctly blocked actions and legitimate actions mistakenly denied by an over-restrictive rule.
Audit trails: evidence of decisions, not proof of perfect safety
The technical note describes OCSF-compatible audit events and optional signatures and hash chaining, intended to make later modification detectable. That can improve provenance and incident investigation. It does not by itself verify that the original event was complete or that all bypass paths were captured. A cryptographically intact log of a misconfigured policy is still a log of a misconfigured policy.
Procurement teams should ask for sample event records with policy version, caller identity, resource scope, decision, exception and timestamp. They should also test log loss, clock differences, key management and whether audit information is segregated from the agent’s authority to rewrite it.
What IBM has not yet demonstrated
Importantly, the article explicitly places some assurances in the future: IBM says it plans to test bypass attempts and investigate combinations of host isolation and network monitoring. This is not a report of completed red-team validation proving that Praxis cannot be bypassed. The absence of a published penetration-test result is not proof that the code is insecure; it is a clear evidence gap for high-assurance use.
Organizations handling health, finance or critical infrastructure operations should demand independent testing in the actual deployment configuration before relying on tool-call controls for sensitive actions. Gateways, permissions and logging form parts of a defense-in-depth strategy; none replaces human authorization for consequential decisions.
A practical black-box acceptance checklist
- Attempt a permitted tool call with the correct user and scoped identity; verify it succeeds.
- Repeat the call with missing or revoked credentials; verify denial and complete audit records.
- Try a disallowed network destination and a restricted field returned in a tool response.
- Test an approval-required action when the approver is unavailable or rejects the request.
- Inject untrusted instructions through a retrieved document and inspect whether policy boundaries still hold.
- Repeat under failure conditions and after a rule update; verify that audit evidence remains traceable.
These tests must be performed in an authorized test environment, never on an unrelated production system. The best measurement is the observed authorization behavior against frozen acceptance criteria, supplemented by evidence of false blocks and operational overhead.
How this fits the enterprise AI stack
We previously described the wider security and operating context in Enterprise AI Agents: Security, Costs and Deployment and Enterprise AI Observability. Praxis is a specific example of policy enforcement infrastructure within that stack, not a substitute for either security governance or monitoring.
DeepTech has reviewed public documentation only. We have not installed PPE, run its test suite or independently demonstrated its efficacy. Its open-source availability and published technical design are observable; claimed protection levels in any particular organization remain to be validated.
