August 2026 Summaries

Researchers at Black Hat 2026 disclosed CoreBreak, a class of flaws affecting AWS Bedrock AgentCore with Strands, Google’s Agent Development Kit, and Vercel’s AI SDK harness that allowed attacker-supplied message histories to trigger tool calls or forge approval events without model-based guardrails being invoked; the vendors issued patches for related CVEs. The article argues that these incidents expose limitations in security controls embedded within an agent’s prompts, model behavior, or message history, because such controls can be bypassed when execution is initiated outside normal model turns. It advocates out-of-band policy enforcement (OBPE), in which tool execution, identity, policy decisions, and approvals are handled through an external typed boundary that agents cannot alter or forge. Redpanda presents its Agentic Data Plane as an implementation of this approach and cites its own experiments across 3,621 trials, reporting substantially lower failure rates for agents protected by OBPE than for prompt-only controls or reviewer models. The discussion also notes that external policy boundaries do not address every risk, particularly general-purpose execution environments with accessible cloud credentials, and concludes that agent systems handling sensitive data should separate authorization and approval mechanisms from the agents they govern.

Redpanda 26.2 introduces expanded Shadowing capabilities intended to simplify migrations from Confluent Cloud, Confluent Platform, and other Kafka-compatible systems by continuously replicating topic data, consumer offsets, ACLs, and Schema Registry contents into a Redpanda cluster. The feature uses API-based Schema Registry synchronization to preserve subjects, schema versions, compatibility settings, references, and identifiers, addressing difficulties associated with manual migrations, MirrorMaker, custom scripts, and disruptive cutovers. Shadowing supports incremental and full synchronization cycles, validation of schemas, configurable handling of unsupported Confluent-specific features, destination write protection during replication, and topic-by-topic failover to enable gradual application migration. Redpanda also supports registry contexts, filtering, and context remapping for complex multi-tenant environments or consolidation of multiple source registries. Beyond migration, the company presents the capability as a way to operate Redpanda as a continuously synchronized disaster-recovery standby for Confluent or Kafka deployments. The release is available for Redpanda Self-Managed 26.2 and Redpanda Cloud BYOC and Dedicated offerings, while API-mode schema replication and Shadowing require an Enterprise license or eligible cloud trial.

Deploying AI agents in enterprise production environments requires stronger governance, control, observability, and data-access restrictions than prototype systems, which are often isolated and lightly secured. Recommended capabilities include immutable, agent-specific logs of inputs, outputs, and tool calls; enforceable guardrails for permissions, budgets, and behavior; rapid shutdown mechanisms; and governance policies maintained separately from agents so they cannot modify them. A control layer between agents and AI models can help condition responses, prevent sensitive-data misuse, monitor token costs, and provide a central point for traffic inspection without impeding development when incorporated into the broader enterprise architecture. Redpanda presents its Agentic Data Plane as such a separate governance layer, offering verified agent identities, tamper-proof replayable audit trails, least-privilege access, spending and activity controls, a unified gateway to models and data sources, and a centralized dashboard for monitoring agents and related services.

Redpanda Operator v26.2 enhances Kubernetes-based Redpanda deployments with multi-region resilience, safer maintenance, and expanded Kubernetes integration. Its generally available Enterprise Stretch Clusters allow one logical Redpanda cluster to span multiple Kubernetes clusters, using synchronous Raft replication to provide stated RPO and RTO of zero for acknowledged data during regional or cloud failures while preserving Kafka API compatibility. The Enterprise beta Pipeline CRD lets teams declaratively deploy and manage Redpanda Connect pipelines with automated cluster credentials, scaling, scheduling, resource controls, and lifecycle status. New community-available per-broker restart probes make rolling upgrades replication-aware by verifying that individual brokers can safely restart and have recovered replicas before subsequent restarts proceed. Gateway API support adds generally available HTTPRoute-based exposure for Redpanda Console and beta TLSRoute-based, SNI-routed Kafka access, offering an alternative to Ingress and enabling shared gateway infrastructure.

The text discusses the challenges and solutions related to the governance of AI agents within enterprises, emphasizing the concept of "agent sprawl" and "shadow AI" due to the ease of deploying AI models without proper oversight. It highlights the limitations of traditional authorization methods, such as borrowing human credentials or using service accounts, which fail to account for the unpredictable nature of AI agents that can bypass or modify their own security measures. The proposed solution is the Agentic Data Plane, a governance framework that enforces policies out-of-band—beyond the agents' reach—ensuring that governance is not merely a suggestion but a robust, unalterable rule. This framework includes features like the Out-of-Band Policy Engine (OBPE), which governs data access and interactions at the server level, independent of individual agents, and allows for fine-grained, attribute-driven permissions. The text also introduces the Agent Network View and forthcoming features like an agent kill switch, emphasizing the need for comprehensive governance that integrates seamlessly with existing systems and operates within the organization's own cloud infrastructure, thereby maintaining control and security over sensitive data. The overall message is that effective governance requires an infrastructure layer specifically designed for managing the unique challenges posed by AI agents, rather than attempting to perfect the agents themselves.

Redpanda SQL has become generally available on Google Cloud Platform (GCP), designed to address the complexities of managing AI agents with advanced database capabilities. Built on the Oxla engine, Redpanda SQL focuses on efficient SQL JOINs and a new ring-buffer-based shuffle algorithm that optimizes core usage, crucial for OLAP systems that face bottlenecks in memory-to-core bandwidth. The platform aims to provide robust agentic governance by enabling real-time analysis and anomaly detection through "agentic kill-switches," which monitor agent behavior and correlations across various data elements. By integrating Postgres compatibility, Redpanda SQL offers a unified architecture that supports open protocols, allowing seamless data integration and governance. The system's design, which includes a policy engine for agent management, facilitates high-performance joins and real-time data processing, enabling enterprises to handle complex agentic workloads effectively. Now deployable in a Bring Your Own Cloud (BYOC) setup, it allows users to maintain data within their own secure environments, ensuring comprehensive control over networking and security.