{"resourceId":"siliconangle-red-hat-agent-security-masked-keys-openshell-preview-20261110","versions":[{"version":"external-4a6a1ee1bffb497b34af983e6cf03fe94b36baa0e391dc4cf4f485b5ecd889bb","resource":{"id":"siliconangle-red-hat-agent-security-masked-keys-openshell-preview-20261110","title":"SiliconANGLE/theCUBE (Red Hat-sponsored): Red Hat layers masked credentials, sandboxes and per-agent identity for AI agents; OpenShell still in preview and identity 'comes next'","organization":"SiliconANGLE Media / theCUBE (interview with Red Hat AI leaders)","sector":"Technology media; vendor interview","geography":"Global; recorded at KubeCon NA, Salt Lake City","publishedAt":"2026-11-10","publicationDate":"2026-11-10","eventDate":"2026-11-10","sourceName":"SiliconANGLE","sourceLabel":"Sponsored video-interview write-up (Red Hat sponsored the segment)","sourceUrl":"https://siliconangle.com/2026/11/10/ai-agent-security-red-hat-layers-masked-keys-sandboxes-kubeconna/","evidenceClass":"vendor-claim","outcomeClass":"emerging","topics":["developers-agents","data-security","infrastructure"],"finding":"Red Hat AI GM Joe Fernandes and chief architect Jessica Forrester describe layered agent security. With OpenShell, agents never receive real keys: credentials are masked on the way in and swapped on the way out. Agents can run in Kata Containers with their own kernels, in full VMs on OpenShift Virtualization, and behind network egress controls. Per-agent identity with least privilege 'comes next', integrating with identity systems. Red Hat is working with OpenAI on OpenClaw Enterprise to manage thousands of personal agents. OpenShell is available only in preview in Red Hat AI, and Fernandes says it is 'approaching' a stable milestone before Red Hat would rely on it for enterprise production.","sledRelevance":"Agencies and universities piloting coding or operations agents need credential isolation and per-agent identity before agents touch eligibility, student or public-safety systems. This source shows those controls are partly preview and partly roadmap, which helps set expectations for public-sector pilots.","evidence":"Opened the full write-up (updated 14:42 EST, 10 Nov 2026), with disclosure that Red Hat sponsored the segment. The claims are statements from executives, with no tests or customers. As a cross-check, I opened Checkable AI's independent Red Hat AI controls page (last checked 4 Nov 2026), which records no supported per-agent identity and OpenShell as Developer Preview.","architectureImplications":"Credential brokering by the sandbox plus VM-level isolation (Kata/OpenShift Virtualization) and egress control form a defense-in-depth stack. Identity-system integration for agents is future work.","governanceImplications":"Agent pilots should record which isolation layer each agent runs in and treat per-agent identity as unavailable in supported form today.","securityPrivacyImplications":"Masked credentials reduce the impact of prompt-injected key leakage if the swap point is trustworthy. Without supported per-agent identity, audit attribution of agent actions may fall back to shared service accounts.","caveats":"Sponsored vendor interview. OpenShell maturity is preview, and identity features are forward-looking. The independent cross-check is from before this edition's window and is used only to confirm current support status.","streamIds":["red-hat"],"roles":{"sales":"Sales framing: Red Hat's agent-security story suits public-sector buyers' caution, but it is partly roadmap. Credential masking and VM-level sandboxing are concrete talking points for agency and university agent pilots. Per-agent identity 'comes next', OpenShell is preview, and an independent controls review found no supported per-agent identity. Position a controlled pilot, not production agents on regulated systems, and let the customer's security office set the exit criteria (interpretation).","engineering":"Engineering framing: evaluate OpenShell as a preview isolation layer. Test whether masked credentials actually stop key exfiltration under prompt injection, and whether egress controls block unapproved destinations. Compare Kata and full-VM isolation against agent workload needs. Because per-agent identity is not yet supported, design interim attribution through dedicated service accounts and audit logs. Revisit the design when Red Hat ships supported agent identity (interpretation).","delivery":"Delivery framing: run agent pilots in sandboxed, non-production environments with explicit exit criteria. Document isolation layer, credential-swap configuration, egress allow-lists and interim identity approach. Acceptance should cover red-team results for credential leakage and blocked egress, with a plan to adopt per-agent identity when Red Hat supports it. Keep production data and regulated systems out of scope until OpenShell reaches a supported release (interpretation)."},"roleDetails":{"sales":{"customerProblem":"SLED teams piloting AI agents need credential isolation and attribution before connecting agents to sensitive systems.","stakeholders":["CISO","Enterprise architect","Research computing or DevOps lead","Identity and access management team"],"discoveryQuestions":["Which systems would agents touch, and with which credentials today?","Can you accept preview-status sandboxing for a pilot?","How would you attribute agent actions without per-agent identity?"],"valueHypothesis":"Bounded: credential masking and sandboxing may reduce damage from compromised agents in pilots.","potentialEngagement":"A sandboxed agent pilot with red-team testing of credential leakage and egress controls.","unsupportedClaims":["Production-ready OpenShell","Supported per-agent identity today","Prevention of all prompt-injection harm"]},"engineering":{"fit":"Good for early agent pilots on OpenShift; not yet a supported production control set.","architectureIntegration":"OpenShell credential brokering plus Kata Containers or OpenShift Virtualization VMs and egress policy.","prerequisites":["OpenShift with sandboxed containers or Virtualization","Red Hat AI with OpenShell preview","Egress allow-list design"],"constraints":["OpenShell is preview","Per-agent identity is roadmap","Sponsored-source claims untested"],"security":"Validate that real secrets never enter the agent context; log all credential swaps and egress attempts.","proofOfValue":["Prompt-injection test attempting key exfiltration","Egress test to unapproved domains","Attribution test of agent actions in audit logs"]},"delivery":{"work":"Stand up a sandboxed agent pilot and red-team its controls.","dependencies":["Preview feature access","Security team participation","Non-production test systems"],"ownership":"Platform engineering with security office co-ownership.","skillsAdoption":"Security staff learn sandbox policy authoring and agent threat modeling.","governanceCheckpoints":["Threat model review","Red-team results review","Decision gate before any production data"],"acceptanceCriteria":["Proposed: zero real credentials observed in agent context during tests","Proposed: all unapproved egress attempts blocked and logged"],"risks":["Preview instability","Weak attribution of agent actions","Over-reliance on roadmap features"]}},"retrievedAt":"2026-11-10T22:12:47-06:00","enrichedAt":"2026-11-10T22:35:00-06:00","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Security teams will need new skills to run sandbox policy, egress rules and future agent identity integration.","procurementImplications":"Do not buy agent platforms on the assumption of supported per-agent identity or GA OpenShell. Require roadmap commitments in writing if those features are essential.","operatingModelImplications":"Define who owns agent sandbox policies (platform vs security team) and how credential-swap rules are reviewed.","sourceVerification":{"openedUrl":"https://siliconangle.com/2026/11/10/ai-agent-security-red-hat-layers-masked-keys-sandboxes-kubeconna/","referenceExcerpt":"OpenShell can also run agents in Kata Containers with their own kernels, inside full virtual machines on OpenShift Virtualization and behind network egress controls","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}