June 2026 Summaries

MCP tunnels provide a method for governing private servers without exposing them to the public internet, addressing the challenge of managing internal servers like Postgres or admin MCPs that are typically hidden behind firewalls. By running a tunnel client within the network that connects to the MCP server, an outbound-only connection to the Speakeasy control plane is established, securely proxying full MCP traffic without opening inbound firewall ports or requiring a public hostname. This solution maintains strong tenant isolation and enhances security by ensuring that only authorized traffic can reach the server, which is now part of the registry with auditable, policy-governed activity. Unlike vendor-specific tunnels that limit connectivity to particular infrastructures, Speakeasy tunnels are vendor-neutral, allowing integration with various agents like Claude, ChatGPT, and the Responses API through a single governed path, effectively creating an internal AI service mesh. This approach facilitates observability and access control for sensitive servers, offering a streamlined, cross-platform connection without compromising security, and is available for organizations to implement through configuration in the Speakeasy dashboard.

Speakeasy has introduced a new feature allowing for per-policy sensitivity tuning in their risk detection system to address the issue of overly broad detection rules that either flag too many false positives or miss genuine threats. Previously, a global confidence threshold applied to all risk policies, limiting flexibility and forcing compromises. The update allows users to set a minimum match confidence score for each policy individually, using a slider in the policy wizard to adjust detection sensitivity according to the specific needs of each policy. This change enables more precise tuning by allowing users to demand higher-confidence matches to reduce noise or accept more false positives to ensure thoroughness, without altering the global default baseline of 0.5. This per-policy sensitivity adjustment is designed to be backward compatible, preventing automatic retuning of existing policies, and can be tested against real data to find the optimal balance between precision and recall.

The post delves into the compliance challenges posed by individual AI licenses in organizations, particularly highlighting the distinction between individual and enterprise AI plans from providers like Anthropic and OpenAI. Individual licenses are economically attractive due to subsidized rates, which offer high-volume AI usage at a lower cost compared to enterprise plans. However, these personal accounts lack the compliance features present in enterprise accounts, such as admin consoles, audit logs, and compliance APIs, creating a "compliance blind spot" that complicates AI governance. To address this, Speakeasy implemented a device agent that records AI usage directly on devices, capturing sessions across both personal and enterprise accounts and generating an identity-attributed, append-only audit log. This solution allows organizations to maintain compliance across all AI usage, even when it occurs on personal licenses, and integrate the data into broader governance frameworks like ISO 42001 and SOC 2.

Speakeasy's new feature for Claude Code sessions provides detailed cost tracking by offering turn-level usage insights, allowing users to see the specific token consumption for each interaction in their chat logs. This feature addresses the challenge of unclear billing by breaking down token usage into input, output, cache reads, and tool calls, enabling users to identify which parts of a session are consuming resources. The system utilizes Claude Code's OpenTelemetry logs to aggregate data per-turn and per-tool, enhancing cost control and observability. With this level of detail, users can differentiate between meaningful work and overhead costs, improving their ability to manage expenses effectively. This capability is seamlessly integrated into existing systems without requiring additional configuration, offering a clearer understanding of resource consumption and facilitating actionable insights for teams operating Claude Code at scale.

In 2026, AI governance has evolved from a secondary concern to a critical infrastructure issue, with projections indicating that spending on AI governance platforms will surpass $1 billion by 2030. The landscape is populated with various vendors offering distinct functionalities under the label of "AI governance platforms," making the choice dependent on the specific job the platform is intended to perform. Five prominent platforms—Speakeasy, Runlayer, MintMCP, TrueFoundry, and Fiddler AI—are frequently evaluated by enterprises, each offering unique strengths in areas such as AI control planes, MCP gateways, model traffic routing, and AI observability. Speakeasy stands out as an AI control plane governing all AI surfaces with inline enforcement and comprehensive audit trails, while Runlayer and MintMCP focus on securing MCP layers, TrueFoundry excels in routing model traffic, and Fiddler AI specializes in AI observability and evaluation. The right choice hinges on whether an organization needs to govern all AI usage, manage model traffic, or evaluate AI behavior, and platforms can often be used in conjunction for comprehensive governance solutions.

In the blog post, Vishal Gowda discusses the challenges and strategies for enforcing AI security policies in real-time agent interactions. The focus is on determining where and how to apply these policies effectively, considering the unique characteristics of each interaction phase: user prompt, tool call, tool response, and model response. The approach involves using a layered detection system, starting with deterministic pattern matching, such as regex, for known patterns, followed by more sophisticated tools like Microsoft's Presidio for PII detection, and finally employing an LLM-as-judge for complex, intent-based policies. This system is designed to prioritize speed and minimize noise, ensuring that security controls remain effective and unobtrusive. The ultimate goal is to develop a dynamic system that can learn and adapt deterministic rules from natural language policies, reducing dependence on costly model evaluations while maintaining robust security.

In addressing the issue of false positives in AI workflows, the focus is on reducing noise by refining policy scope rather than compromising detector accuracy. This involves narrowing the policy scope with specific conditions such as individual server or tool functions and employing role-based access control (RBAC) to target specific users or groups, thereby preventing unnecessary alerts. Exemptions are used to avoid counting certain interactions as violations, either by preventing policy evaluation or suppressing findings post-evaluation. Built-in detection rules are provided for initial setup and can be tested against real data to fine-tune effectiveness without weakening the policy. These strategies form part of a broader enforcement model that includes synchronous inspection of interactions within the AI control framework.

AI guardrails are essential for safeguarding against prompt injection and tool poisoning attacks, which have evolved from theoretical demonstrations to real incident concerns. To effectively mitigate these risks, Speakeasy offers AI guardrails that enforce deterministic and auditable risk policies across all prompts and tool calls within a company. These policies include scanning for prompt injection attempts, flagging destructive commands, and implementing prompt-based policies, all of which are tailored to specific users or roles to prevent obstructions. The system provides transparency by offering custom messages when a call is blocked, allowing users to understand the reason and take corrective action. Speakeasy's approach is designed to be adaptable, starting with a read-only posture to monitor risks before tightening policies, and offers additional support for teams needing guidance in setting up their policy framework.

Speakeasy provides a streamlined solution for managing MCP servers, allowing users to add both self-hosted and remotely hosted servers from SaaS vendors into a unified MCP registry with consistent governance, security, and monitoring features. By simply entering a server's URL, Speakeasy checks its compatibility through an MCP initialize request, and once verified, the server integrates into the registry with the same access controls and audit trails as those built on Speakeasy. This integration ensures that network-level security policies are uniformly applied, and all server activities are observable within a single dashboard, enhancing trust and oversight by eliminating fragmented monitoring systems. This comprehensive approach simplifies the management of MCP servers, offering a cohesive view of server operations and usage patterns under one roof.

Speakeasy's shadow MCP detection and governance system provides a strategic approach to managing AI connectivity, addressing security risks associated with unauthorized server usage. As traditional bans prove ineffective, this system detects shadow AI usage, evaluates server risks, and facilitates access for users who need it, thus aligning security measures with team workflows. Speakeasy employs runtime hooks to identify and record connections to unsanctioned servers, allowing organizations to create explicit access rules based on server identity, URL, or host, with a precedence for denial in conflicting cases. This proactive governance approach includes a user-friendly request-access feature, which enables developers to seek approval for blocked connections, thereby reducing friction and potential workarounds. By offering a dashboard for overseeing shadow MCP usage, Speakeasy ensures a controlled yet flexible environment for AI server interactions, enhancing both productivity and security.

The Speakeasy Team's update introduces enhancements to their product by tying agent actions to real identities, enabling organizations to trace every allow-or-deny decision back to an individual, thus addressing the crucial question of "who did this and why were they allowed to?" The system integrates with identity providers through directory sync, automatically updating roles and access based on changes in user attributes, ensuring accurate and current access control. The update also refines access decision-making with deny-wins semantics, ensuring safe outcomes even in complex rule sets, and introduces authorization challenge logging for retrospective analysis. This logging feature allows organizations to understand past access decisions and make informed adjustments, supported by identity-stamped telemetry and directory sync which are now available, with challenge logging available on a per-organization basis. Organizations can configure and manage these features through the WorkOS admin portal to gain comprehensive insights into access control across their systems.

Speakeasy introduces Remote MCP, a solution designed to manage and govern access to multiple MCP servers, whether self-hosted or provided by third-party vendors. By creating an MCP endpoint that proxies any server via its URL, Speakeasy allows teams to enforce governance and manage access controls, ensuring all tool calls are logged and monitored within a unified dashboard. The system supports various authentication methods, including OAuth, managed credentials, and header pass-through, which streamline the integration process while preserving security and user identity. This approach enables organizations to maintain control over their server interactions without requiring changes to the upstream servers themselves, even for servers behind firewalls, by leveraging outbound-only connections for secure access.

The comparison between Speakeasy and Workato Enterprise MCP highlights their distinct approaches in the field of enterprise AI governance. Speakeasy focuses on security and governance by providing an AI control plane that inspects traffic, blocks threats in real-time, and measures AI adoption across organizations. Its architecture allows for governance of a wide range of servers, including third-party and self-hosted ones, and offers detailed analytics on AI usage. On the other hand, Workato, an integration company, leverages its existing iPaaS framework to transform recipes, connectors, and API collections into MCP servers. Its security model centers on identity and access management with features like Verified User Access, which allows agents to operate with user-specific permissions. Workato's platform is advantageous for organizations already utilizing its services, as it efficiently integrates existing automations into MCP tools. The choice between the two depends on whether an organization prioritizes comprehensive security for all AI interactions or seeks to leverage existing Workato infrastructure for MCP functionalities.

In the wake of the leaked details about Anthropic's Claude Mythos, the cybersecurity landscape experienced significant turmoil, with market reactions reflecting fears of AI's potential to favor attackers. Despite this, the same AI models have demonstrated substantial defensive capabilities, identifying over 10,000 high-severity vulnerabilities in a short period, although this success has been underreported compared to offensive narratives. The Mythos model, which has proven adept at autonomously finding and exploiting vulnerabilities, underscores the importance of robust internal governance over AI systems. Enterprises are urged to implement AI Control Planes to regulate and secure AI interactions, ensuring agents operate within safe boundaries. This proactive governance is crucial as AI capabilities continue to advance, with the potential for both significant risk and protection, depending on how effectively these tools are managed.

The era of subsidized AI is ending as leading AI labs like Anthropic and OpenAI prepare for public markets, leading to significant price increases for AI models and services. Enterprises that previously benefited from flat-rate plans are now facing usage-based billing, highlighting inefficiencies in AI spending. As AI spending becomes more transparent, companies must address infrastructure challenges to efficiently manage costs. Despite falling per-token prices, overall AI bills are rising due to increased demand for frontier models, higher token consumption in agentic workloads, and the shift from flat-rate to metered pricing. The emerging discipline of FinOps for AI, which includes attributing costs, routing tasks to cost-effective models, and optimizing tool usage, aims to manage these expenses effectively. An AI Control Plane, serving as a governing layer between agents, models, and tools, helps organizations track and reduce AI costs by implementing strategic controls and ensuring spend attribution.

Speakeasy and Fiddler AI are both positioned as AI control planes, but they serve different purposes within enterprise AI governance and observability. Speakeasy functions as a governance-focused security platform that serves as an enforcement layer, providing real-time threat blocking, policy enforcement, and audit logging by sitting directly in the traffic path. This allows it to block policy violations immediately and manage access control at the tool level. In contrast, Fiddler AI is primarily an observability and evaluation platform that monitors AI model performance, providing insights into drift, hallucination, and safety scoring without directly enforcing policies in the traffic path. Instead, Fiddler's platform focuses on agent tracing, risk evaluation, and compliance reporting, relying on application logic to handle blocking based on its scoring API. Both platforms achieve different objectives in managing AI systems, with Speakeasy prioritizing access governance and Fiddler focusing on evaluating AI model quality. They can be used together to provide comprehensive AI system management, combining Speakeasy's direct enforcement with Fiddler's detailed evaluation capabilities.

Speakeasy and TrueFoundry are two distinct platforms aimed at addressing different aspects of AI integration within enterprises. TrueFoundry primarily functions as an AI gateway company, focusing on efficiently routing application traffic across a vast array of over 250 large language models (LLMs), emphasizing features like load balancing, failover, and cost control, especially beneficial for teams managing model traffic and self-hosted models on Kubernetes. In contrast, Speakeasy is designed for AI security and governance, providing a comprehensive control plane that includes native threat detection, policy enforcement, and detailed audit logging across all AI interactions within an organization, from employees to AI clients, which do not pass through a server-side gateway. While TrueFoundry excels in model traffic management with features like multi-provider routing and self-hosted deployment options, Speakeasy provides deep governance capabilities with built-in security measures and governance workflows, catering to organizations prioritizing secure and measurable AI usage.

Nolan Di Mare Sullivan discusses the implementation of tag-based tool filtering for Speakeasy-hosted MCP (Multi-Channel Platform) servers, highlighting its necessity due to Datadog's extensive tool list that can overwhelm an AI agent's context window. Datadog's MCP server, which exposes 142 tools across 22 toolsets, demonstrates the need for efficient tool filtering to improve agent performance by reducing noise and focusing on relevant tools. Tag-based tool filtering allows server operators to control which tools are visible to clients by tagging tools at the source, thus optimizing the context window and enhancing tool selection. This approach contrasts with dynamic toolsets, which optimize tool loading at runtime, and both methods together offer a comprehensive solution for managing large MCP servers. The article emphasizes that this filtering is a non-breaking change, as it is opt-in and maintains the full tool list when no tags are specified, ensuring backward compatibility and flexibility for server operators.

Speakeasy and Runlayer are AI governance platforms designed to secure and manage AI usage within organizations, but they differ significantly in their approach and scope. Speakeasy serves as a comprehensive AI control plane that governs the entire path between AI systems and organizational infrastructure, including model calls, MCP tool calls, agents, and assistants, while providing features like observability, cost-tracking, and policy enforcement. It excels in generating governed connectors from internal API contracts and offers deployment flexibility without the need for MDM. In contrast, Runlayer focuses on MCP enablement, securing and managing tool usage through a comprehensive policy engine, real-time tool scanning, and endpoint reach for shadow AI detection using MDM-deployed agents. While both platforms offer robust security and governance features, Speakeasy is better suited for organizations seeking governance across all AI surfaces with ISO 27001 certification, whereas Runlayer is ideal for MCP-centric programs prioritizing device-level control and existing MDM infrastructure.

Speakeasy and MintMCP are two AI governance platforms designed to secure and manage AI usage within organizations, but they differ significantly in their focus and capabilities. Speakeasy serves as an all-encompassing AI control plane, overseeing the entire AI usage path, including model calls, MCP tool calls, agents, and assistants, and it emphasizes end-to-end governance and enablement, offering features such as observability, cost-tracking, and policy enforcement. MintMCP, on the other hand, is primarily an enterprise MCP gateway and agent monitor that secures the Model Context Protocol (MCP) layer, providing role-based access, policy enforcement, and audit logging specifically for MCP and coding agents. While both platforms offer managed connectors and compliance features like SOC 2 and HIPAA, Speakeasy extends its compliance with ISO 27001 and GDPR and has a longer track record in production environments. The choice between the two largely depends on whether an organization seeks comprehensive AI governance with more deployment flexibility, as offered by Speakeasy, or a focused MCP enablement solution with managed hosting and agent monitoring, as provided by MintMCP.

Speakeasy has introduced a network-level access feature for its MCP servers, allowing organizations to restrict server access to trusted networks, enhancing security by limiting exposure to internal users only. This is achieved by setting an IP allowlist for custom domains, ensuring that only traffic from specified IPv4 addresses or CIDR ranges can reach the servers, while unwanted traffic is blocked at the ingress level. This approach leverages Kubernetes networking, applying the allowlist to both NGINX ingress and Envoy gateway paths, ensuring requests from outside the allowed network are refused with a 403 error before reaching the server. Organizations can configure network-level access at the domain settings, ensuring private servers remain accessible while maintaining security with no bypassing of the policy. This feature represents the first tier of network-level security for MCP, with further integrations planned, catering to the increasing demands of security teams.

The EU AI Act mandates stringent governance for high-risk AI systems, making it illegal not to have an AI Control Plane, a governing infrastructure that ensures compliance through connection, identity, policy enforcement, and observability. This legislation categorizes AI systems by risk, with high-risk applications, such as those used in employment, credit scoring, and essential services, requiring automatic record-keeping, human oversight, continuous risk management, and data governance. Non-compliance can result in substantial fines, similar to GDPR penalties. The Act's deadlines have been adjusted, with obligations for high-risk systems taking effect in December 2027 and August 2028. The AI Control Plane, which acts as a unifying layer for managing AI interactions, is critical for meeting these requirements, providing traceability, policy enforcement, data governance, and a controlled environment for AI operations. Enterprises like Uber and JPMorgan Chase have already implemented such infrastructure, treating it as a competitive advantage, while others are urged to adopt this technology before the compliance deadlines.