By submitting this form, you agree to our 

 is right around the corner, and there are 11 clear new Python Enhancement Proposals (PEPs) added to the Python project! There are several over significant changes listed within the 

Python 3.14 brings a whole bunch of useful build improvements, including discontinuation of PGP signatures in 

. Python versions 3.14 and onwards will no longer provide PGP signatures for release artifacts. Instead, Sigstore is recommended for verifiers. Similarly, free-threaded mode (

), which was introduced in the previous version, 3.13, has undergone major enhancements. The implementations outlined in this enhancement are now complete, incorporating updates to the C API, and the interpreter’s earlier temporary fixes have been replaced with long-term solutions.

The Cloudsmith team is really excited about this release and everything that comes with it! Note that some of the statuses may change between the time of this blog being posted and the final release date. Knowing that, let’s jump into all of the new features and improvements in Python 3.14.

Here are a few of the changes that Cloudsmith employees are most excited about in this release:

PEP 761: Discontinuation of PGP signatures

“From my perspective at Cloudsmith, the move away from PGP signatures in Python is a welcome change. PGP has long struggled with usability and the burden of managing long-lived private keys, and while it was the default for artifact signing, it never felt ideal. Sigstore offers a much more practical approach, using short-lived keys tied to human-readable identities, and it’s already gaining traction across ecosystems like PyPI, NPM, Homebrew, and GitHub. We’re also seeing projects we work with, like 

, embracing Sigstore, which reinforces that this shift is a real step forward for security and usability.”

PEP 734: Multiple interpreters in the stdlib

“This was an inevitable development, in my opinion. While the CPython runtime has supported running multiple interpreters, which are independent copies of Python within the same process, for well over 20 years at this point, the feature was still limited to the C-API and thus saw little adoption due to low awareness and the absence of a standard library module. With Python 3.14 introducing the new 

 module, this long-standing capability is finally becoming more accessible, marking a significant step toward mainstream usage. While current limitations remain, their impact will diminish as CPython evolves and the community contributes solutions via 

PEP 784: Adding Zstandard to the standard library

“Python 3.14’s addition of Zstandard (via 

) is an exciting upgrade because it brings one of the fastest, most efficient modern compression algorithms directly into the standard library, alongside a new unified 

, consistently outperforms older formats like 

 by offering both higher compression ratios and dramatically faster decompression speeds, which has made it the de facto standard across filesystems, packaging tools, and industry at large.

By including it natively, Python eliminates the confusion of multiple third-party bindings, enables direct support in core modules like 

, and opens the door for faster, smaller Python wheels and other packaging improvements. The reserved 

 namespace not only avoids naming conflicts with PyPI packages but also creates a clean, future-proof home for compression libraries, much like 

 does for hashing. This change reflects Python’s “batteries included” philosophy while giving developers a state-of-the-art compression tool that is both practical today and well-positioned for long-term usage.”

Free-threaded Python is now officially supported

 this enhancement proposal, which actually sets requirements for advancing 

, which was the larger, ongoing effort to make Python’s 

) optional. PEP 703 outlines a three-phase plan:

First phase introduces an experimental GIL-free build in Python 3.13

Phase two makes this build officially supported - though still optional

While Phase 3 makes it the default. 
This new feature defines the expectations for the second phase, ensuring the free-threaded build is stable, usable, and performant enough for broader adoption.

The motivation behind this shift is to balance the benefits of free-threaded Python, which offers greater concurrency, reduced latency, and expanded threading capabilities against costs such as performance penalties, memory overhead, and ecosystem complexity. Current benchmarks show around a 10% slowdown and 15-20% more memory use compared to the traditional GIL build, both considered acceptable trade-offs for Phase 2.

The APIs introduced so far have proven stable, and further refinements are expected without breaking changes. With growing third-party support and clear criteria for performance, memory use, documentation, and API stability, Python 3.14 is targeted as the point where free-threaded Python becomes an officially supported build.

Standard Library Support for Isolated Python Interpreters

Python 3.14 sees the addition of a new standard library module, 

, to expose CPython’s long-standing support for multiple interpreters in a single process. Each interpreter is strictly isolated, with its own modules, classes, and variables, while sharing only certain immutable or low-level process state. Unlike threads, which share most of the same runtime state, interpreters provide stronger isolation, making them better suited for parallelism, especially now that 

The new module will provide a high-level interface for creating, inspecting, and executing code in interpreters, as well as managing communication through a built-in, cross-interpreter-safe 

. This brings powerful functionality, previously limited to the C API, into Python code. An additional 

concurrent.futures.InterpreterPoolExecutor

 will be provided to simplify concurrent workloads.

The interpreters module defines an Interpreter object with methods like 

 to initialise globals. Interpreters can safely exchange data using 

, which supports pickle-based transfers and special handling for buffer objects like 

. This queue acts as a synchronisation and communication mechanism, enabling patterns like worker coordination across interpreters. For example, the below snippet taken from the PEP documentation shows multiple interpreters sharing a 

 of large data, synchronised with a queue token:

Making Parentheses Optional in Except and Except Blocks

 blocks in Python, users can now catch multiple exception types, so long as the “

” operator clause is not used. Currently, parentheses are mandatory around multiple exception types, a requirement carried over from Python2.

Removing this requirement simplifies the syntax, improves readability, and makes it more consistent with other comma-separated constructs in Python, such as function arguments and tuple literals. Parentheses will still be required when using as to capture the exception instance to avoid ambiguity. The change is purely syntactic, fully backwards compatible, and does not alter exception handling semantics.

Deferred Evaluation Of Annotations Using Descriptors

Python annotations let developers attach type information and metadata to functions, classes, and modules. Traditionally, annotations were evaluated eagerly when the object was defined, which led to common problems with forward references and circular dependencies. 

 attempted to fix this by “stringizing” annotations (storing them as strings), solving the forward-reference issue but breaking runtime use cases where actual objects were expected instead of strings.

A new proposal introduces a third, more flexible approach: lazily evaluating annotations through a dedicated 

 method. At compile time, Python generates a helper function that constructs the annotation dictionary, then attaches it to the object as 

. Accessing these annotations triggers this function, evaluates the annotations only when needed, and caches the results. This design eliminates circular-reference issues, preserves runtime usability, and allows for new introspection capabilities. Furthermore, 

 gain a new format parameter, letting users choose between evaluated values, forward-reference proxies, or source-code strings for annotations.

Here’s a simplified illustration of how this lazy evaluation model works:

With this change, annotations retain their runtime meaning while still avoiding forward-reference pitfalls, which supersedes the enhancement #563 string-based approach mentioned earlier, as well as offering a unified solution for both static and runtime use cases.

Safe external debugger interface for CPython

This introduction of a zero-overhead debugging interface for CPython should allow debuggers and profilers to safely attach to running Python processes. By providing well-defined safe execution points within the interpreter’s evaluation loop, external tools can request code execution without modifying normal execution or introducing runtime overhead.

This mechanism enables scenarios like attaching pdb to live processes by PID, similar to 

, allowing developers to inspect, evaluate, and step through running Python code interactively without stopping or restarting applications. Current approaches to live debugging in Python are risky because they rely on unsafe code injection, which can execute at arbitrary points in the interpreter, leading to crashes, memory corruption, or deadlocks. By integrating a small support structure into each thread state and extending the debug offsets table, this proposal ensures that debugging code is executed only when the interpreter is in a consistent and safe state.

The interface works by having debuggers write control information, such as the path to a Python script, into dedicated memory locations in the target process. The interpreter checks for pending debugger requests at existing safe points (via the 

), executing the provided code without interrupting critical operations like memory allocation or garbage collection.

 API simplifies remote execution for external tools, while environment variables and build flags allow redistributors and administrators to disable the feature if desired. Multi-threaded considerations are handled naturally: injected code runs under the normal GIL, ensuring thread safety, and can target individual threads or all threads as needed. This approach aligns Python with other major languages in live debugging capability while maintaining safety and zero runtime overhead during normal execution.

Deprecating PGP signatures for CPython artifacts

Again, this PEP isn’t merely a deprecation announcement. Instead, we see the introduction of an entirely new security model through Sigstore. Since 

, CPython artifacts have been signed with both PGP and Sigstore. PGP relies on long-lived private keys, which release managers must maintain and protect for many years, creating ongoing risks and burdens. In contrast, Sigstore uses short-lived keys tied to human-readable identities via OpenID Connect, greatly simplifying the signing process while providing stronger ergonomics. Sigstore is already gaining wide adoption across ecosystems like PyPI, NPM, Homebrew, and GitHub.

This new feature will not only deprecate, but will eventually discontinue the PGP signatures for CPython altogether, transitioning fully to Sigstore. Current stable releases (3.11 through 3.13) will continue offering PGP signatures until end-of-life, but new release managers (starting with Python 3.14) will not be required to generate them. Maintaining dual systems has slowed adoption of newer methods, creating a “

” that discourages verifiers from moving forward.

By dropping PGP, CPython can encourage downstream ecosystems to fully adopt Sigstore, while still providing flexibility for extraordinary delays if needed. Though this change shifts reliance to Sigstore’s centralised infrastructure, it aligns with the security practices already expected of release managers and leverages services better suited for long-term key management. Documentation will guide integrators, such as Linux distros and package managers, on verifying artifacts with Sigstore.

Disallow return/break/continue that exit a finally block

Nigel Douglas

Head of Developer Relations

Python 3.14 – What you need to know 

 launched with the goal of giving developers a way to verify the integrity and publisher of container images. It was built on The Update Framework (

Fast forward to 2025: Notary is no longer actively maintained, the industry has evolved, and 

 of Docker Hub image pulls use DCT. Microsoft is dropping support in Azure Container Registry, and now Docker is formally beginning to retire DCT. This starts with Docker Official Images (

If you're using DCT in your pipelines, things may already be breaking.

But there’s good news: the ecosystem is moving forward through 

, an open standard built for the modern software supply chain.

Provenance is becoming a critical part of how we secure the modern software supply chain. At its core, it’s about traceability: being able to answer where a software artifact came from, how it was built, and who was responsible at each stage of the process. Threat actors continue to increasingly target dependencies, build systems, and distribution infrastructure, meaning verifiable metadata about your software’s origin is really important. Being able to establish the provenance of your container images helps teams make informed, trustworthy decisions about what they built, run, and deploy. It’s a key building block for adopting frameworks like 

, and is top of mind for teams in highly regulated environments.

But knowing what happened during a build isn’t enough, you also need a way to attest to that information in a way that others can verify, without adding unwanted friction. So when you think, “why do I need Sigstore?” - that’s the reason. Sigstore is an open-source project that enables cryptographic signing of software artifacts like container images using standard tooling like Cosign, Fulcio, and Rekor.

Unlike DCT, which relied on TUF until the upstream Notary codebase was no longer actively maintained, Sigstore has an active community of contributors and is integrated into a long list of 

Sigstore makes it easy to integrate signing and verification into your CI/CD pipeline without the need for custom certificate infrastructure requirements. You can use keyless signing (via OIDC), or generate and manage your own keys.

While both Sigstore and TUF aim to improve the same software supply chain security issues, they address different layers of the trust and verification problem. For instance, TUF focuses on securing software update systems. It defines certain roles (such as root, targets, and snapshots) as well as providing a metadata-driven approach to protect against rollback attacks or mix-and-match attacks. TUF’s strength and differentiation is in its resilience, so even when some keys are compromised, it can still recover securely.

Whereas, Sigstore is designed specifically for artifact signing and transparency, especially within CI/CD-related contexts. Likewise, Sigstore offers a developer-friendly approach to cryptographic signing, as stated already, it leverages OIDC-identity (keyless) signing, open and auditable transparency logs through Rekor, and dedicated integration with Fulcio for certificate issuance.

Fundamentally, these two approaches have key distinctions in that Sigstore ties identities to signatures (like Github Identity), whereas TUF focuses only on robust update metadata and key delegation. If you’d like to learn more about how these two solutions differ, Bob Callaway and Dan Lorenc delivered a

 on the topic of why Sigstore was created.

Again, Cloudsmith already has native support for Sigstore's Cosign. When you enable “cosign signing" for your repository, every Docker image uploaded to a Cloudsmith repository is 

An ECDSA private key is generated when you create a repository.

Images are signed automatically on upload using this key.

You can download the matching public key to verify signatures.

Or you can use your own Cosign keypair and upload external signatures via the 

This means Cloudsmith customers already have a modern, secure signing solution in place without the need to scramble around for a replacement DCT.

Migrating from Docker Content Trust to Sigstore (Cosign)

 and TUF, which allowed for signature-based verification of container images. But with DCT now deprecated and official support ending, it's time to transition to modern, open tooling.

Here is how to migrate to Cosign with Cloudsmith:

 set in your environment, unset it. Continuing to use DCT after August 8, 2025, will cause failures when pulling Docker Official Images. Unsetting the below environment variable will avoid any unwanted pull failures:

To verify a signed image from Cloudsmith:

 feature flag tells Cosign not to query the Rekor transparency log, which is appropriate since Cloudsmith-hosted signatures are private by default.

In scenarios where you’re using a default Cloudsmith-managed keypair, you’ll need to download the associated ECDSA public key (making sure to replace 

 with your actual Cloudsmith workspace and repository names):

3. Sign images yourself (totally optional)

Prefer signing artifacts manually? You can generate and upload Cosign signatures directly:

Then upload both the image and signature to Cloudsmith. We’ll store and associate them automatically.

Cloudsmith does not log signatures to Rekor when signing on upload, because most customers use private repositories. As a result:

Cosign will warn if you attempt to verify without setting 

To avoid confusion, always include the flag when verifying Cloudsmith-hosted signatures:

f you want to use Rekor for transparency, you can opt into it by signing manually with Cosign and uploading both the signature and the image to Cloudsmith.

. By migrating from Docker Content Trust to Sigstore (via Cosign), you're adopting an open, scalable approach to supply chain security, while simultaneously staying compliant and adhering to industry best practices. And with Cloudsmith, this migration is already built in.

Migrating from Docker Content Trust to Sigstore

 is right around the corner, and there are quite a lot of changes to unpack! Removing enhancements with the status of “Deferred” or “Removed from Milestone” we have 59 Enhancements in all listed within the 

Kubernetes 1.34 brings a whole bunch of useful enhancements, including 45 changes tracked as ‘

’ in this Kubernetes release. From these, 

 are graduating to stable, including support for Direct Service Return (

) and overlay networking in Windows kube-proxy.

 features are also listed in the enhancements tracker, including a built-in capability for pods to authenticate to kube-apiserver using mTLS. This is especially cool as it's built from pieces that are easy for external projects to reuse to deliver additional functionality using X.509 certificates. Amongst other new benefits, DevOps teams will benefit from new container restart rules. These rules declare the container-specific logic that can be used to customise the restart behaviour of the container when it exited or crashed.

If you don’t have time to read the entire release report right now, Gerome Grignon provides an easy to consume, status overview of each feature release from 1.19 through to 1.34 - 

. The Cloudsmith team is really excited about this release and everything that comes with it! Let’s jump into all of the major features and changes in Kubernetes 1.34.

“From a personal perspective, I find the introduction of CEL-based mutating admission policies in Kubernetes 1.34 incredibly exciting because it brings clarity and simplicity to an area that’s historically been complex and operationally heavy. For those dealing with the overhead of managing external mutating webhooks (whether for debugging latency issues, ensuring HA, and handling failure edge cases) this feels like a major leap forward. Being able to express common mutations declaratively using CEL, directly within the API server, eliminates an entire layer of moving parts while still retaining flexibility. It also aligns beautifully with the Kubernetes philosophy of declarative configuration, and I’m particularly excited about how this could streamline GitOps workflows and simplify secure, consistent policy enforcement across environments.”

Nigel Douglas - Head of Developer Relations

#4427 Relaxed DNS search string validation

“I'm really happy to see this KEP graduate to stable in Kubernetes 1.34 because it finally addresses a long-standing pain point when dealing with legacy or standards-bending DNS setups. In the past, trying to use valid-but-practically-common patterns like underscores or a single dot in dnsConfig.searches meant hitting a frustrating wall with 

 validation. This often forced awkward workarounds or even blocked legacy workloads from being migrated into Kubernetes entirely. With this KEP now stable, Kubernetes is acknowledging real-world usage patterns while still maintaining safe, opt-in behaviour.

From a practical standpoint, it's empowering for me to be able to configure a pod with dnsConfig.searches like abc_d.example.com without having to fight the platform. These are small changes, but they remove major friction, especially when you're dealing with SRV records or short-name resolution in systems not designed with strict DNS compliance in mind. The design is also thoughtful: it's gated, it respects compatibility boundaries, and it won't break clusters on upgrade or downgrade. To me, this shows Kubernetes is maturing in a way that balances standards compliance with usability, and that's worth celebrating.”

Esteban Garcia - Principal Software Engineer

#4639 VolumeSource: OCI Artifact and/or Image

“This is a game-changer. Allowing OCI images to be mounted as volumes is incredibly exciting for us. It means our customers can now leverage their existing OCI registries to seamlessly distribute ML models, configurations, and binaries directly to their pods. This move greatly simplifies artifact management, streamlines MLOps and CI/CD pipelines, and enhances security. All of it while using familiar, standard tooling. I can’t wait to document some new examples about this.”

Pablo Javier López Zaldívar - Technical Writer

Allow recreation of pods once fully terminated in job controller

 require dedicated pods per index, but issues arise when these pods are prematurely terminated due to factors like preemption or eviction. Currently, replacement pods are created immediately, which often results in failure to start - especially in clusters with limited resources or strict budgets. This premature creation can delay scheduling, as resources may not be available until the terminating pod has fully exited. Additionally, if cluster autoscaling is enabled, these early replacements may trigger unnecessary scale-ups.

To address this, the proposed change introduces more control and visibility into pod replacement behaviour for Kubernetes Jobs. The Job controller will now differentiate between active, failed, and terminating pods. A new 

 field will track the number of terminating pods, and an optional spec field will allow users to configure whether replacement pods should wait for termination to complete. This information can also support queueing controllers like Kueue in better quota calculation. The goal is to offer more flexible and efficient scheduling without affecting other workload APIs.

This enhancement introduces support for a 

. Currently, StatefulSets update their pods one at a time, which can be slow and inefficient for certain workloads. By allowing multiple pods to be updated simultaneously (up to the specified 

 value). This change enables faster and more flexible rolling updates for stateful applications.

The motivation stems from several real-world scenarios. For instance, StatefulSets preserve pod identities, making them ideal for applications that publish consistent metrics or rely on fixed pod names - something not possible with Deployments. Additionally, applications that perform lengthy startup tasks or can tolerate multiple replicas going down (e.g., follower nodes in a distributed system) would benefit from the faster updates enabled by 

 for tracking updates, offering simpler and more predictable revision management compared to Deployments. The proposal aims to make StatefulSets more feature-rich and deployment-friendly, while retaining their core identity-preserving capabilities.

Deployments in Kubernetes currently show inconsistent behaviour when dealing with terminating pods, depending on the rollout strategy or scaling activity. In some cases, it might be more efficient to wait for pods to fully terminate before creating new ones, while in others, launching new pods immediately is preferable. To address this inconsistency, a new field, 

, is proposed to give users control over when replacement pods should be created during deployment updates.

The goal is to let users define whether a Deployment should wait for pods to terminate before spinning up new ones or proceed without delay, all while respecting the chosen deployment strategy. Additionally, the status of Deployments and ReplicaSets would be enhanced to reflect the number of managed terminating pods. This distinction is important, as pods marked for deletion (with a 

 field. This can lead to temporary over-provisioning - especially during rollouts or unexpected deletions like evictions - causing resource strain, potential scheduling issues, and even unnecessary autoscaling, which may increase cloud costs in large-scale or tightly resourced environments.


This proposal outlines a plan to enhance Kubernetes scalability and performance by enabling consistent reads directly from the watch cache instead of querying etcd. The core idea is to ensure that cached data is up-to-date by leveraging etcd's progress events, which indicate the latest revision observed and guarantee that all future events will be newer. This mechanism allows Kubernetes to verify how fresh the watch cache is and safely serve consistent (

) reads from it, without needing to go back to etcd each time.

The motivation behind this change is significant: serving reads from the cache is far more efficient than querying etcd, filtering, deserialising, and garbage-collecting large numbers of objects. This is particularly impactful for node-originated requests, such as kubelets listing pods scheduled on their nodes. Instead of listing and filtering hundreds of thousands of pods from etcd, the cache can immediately return just the relevant data. Beyond performance, this also resolves the long-standing “

” issue, where reflectors could receive outdated data on startup due to defaulting to non-consistent reads.

The implementation plan includes using a consistent read to get the latest resource version, waiting for the watch cache to reach that revision (with repeated 

 calls every 100ms if necessary), and then serving the request once the cache is fresh enough. This approach introduces little to no extra scalability cost and enables Kubernetes to maintain correctness while improving efficiency - particularly in large-scale clusters.

 as a simpler, declarative alternative to 

 in Kubernetes. Many common mutation scenarios (such as adding labels or injecting containers) can now be expressed using concise 

 (Common Expression Language) expressions, significantly reducing operational complexity.

This method eliminates the need for running and managing external webhooks and provides better introspection and performance. Because these policies run in-process within the Kubernetes API server, they avoid the latency and reliability issues associated with webhooks and support automatic re-invocation when needed.

, depending on the desired semantics and compatibility needs. The structured merge strategy is preferred for its declarative nature and integration with 

A key feature is the use of CEL expressions to dynamically construct the partial object to be merged into the original request. For example, injecting a sidecar initContainer can be done using the following 

This enhancement provides a Kubernetes-native way to express and apply mutations with clarity, speed, and declarative consistency - without sacrificing flexibility. It also paves the way for policy frameworks to be built around in-tree mechanisms and reused across systems like GitOps or CI/CD, all without relying on external webhook infrastructure.

This proposal introduces a new approach to handling leader election transitions for key Kubernetes components (

) by allowing them to gracefully release their leadership lock and revert to a follower state without restarting the entire process. Currently, when a component loses leadership, it forcefully exits using 

, requiring the kubelet to restart it. This results in unnecessary overhead and prevents clean shutdowns. The proposed change, gated behind a new feature flag (

), aims to improve high-availability (HA) behaviour by enabling components to stop their internal controllers, release resources, and then attempt to reacquire the lease - all without exiting.

To implement this, the components' control loops will be updated to support proper context cancellation and cleanup when leadership is lost. Key technical changes include modifying the 

 callback, enhancing health check de-registration, and deferring or resetting resources upon leadership loss. While this offers clear benefits in reduced latency and better lifecycle handling, there are associated risks - such as memory leaks, failure to respect shutdown signals, and regressions from future code changes. These risks are mitigated through audits, new tests, best practice documentation, and conditional gating of shutdown logic.

This enhancement proposal introduces KYAML, a new output format for kubectl (invoked via 

 is a strict subset (or dialect) of YAML designed to be compatible with existing YAML tooling while avoiding many of YAML’s most common pitfalls. It emphasises syntactical rules that prevent ambiguous behaviour, such as always quoting strings to avoid unintentional type coercion (for example "

" being interpreted as a boolean), using explicit flow-style syntax (

 for maps), and eliminating reliance on whitespace sensitivity. These design choices make KYAML easier to read, write, patch, and template, particularly beneficial in tools like Helm.

The KEP also proposes making KYAML the standard format for all Kubernetes project-owned documentation and examples. The motivation stems from YAML’s complexity and its tendency to produce valid but incorrectly interpreted files due to issues like indentation or ambiguous values. While JSON is an alternative, it lacks comments and is less human-friendly. KYAML aims to blend the safety and clarity of JSON with YAML’s readability and flexibility. By updating examples and tooling to use KYAML, the project hopes to guide the ecosystem toward more reliable and error-resistant configuration practices.

 Implement a .kuberc file to separate user preferences from cluster configs

 phase, this feature will be enabled through the 

. Users can override this location using the 

). The kuberc file will serve as an optional configuration to separate user preferences from cluster credentials and server configurations. This proposal aims to introduce a new file dedicated to user preferences, which will offer better flexibility compared to 

, which currently under-utilises its preferences field due to the creation of new files for each cluster that mix credentials with preferences.

The kuberc file will provide a clear separation between server configurations and user preferences, allowing users to define command aliases, default flags, and other customisations. The file will be versioned to support easy future updates, and users will be able to maintain a single preferences file, regardless of the 

 environment variable. This proposal also suggests deprecating the kubeconfig preferences field to streamline configuration management. The kuberc file will be entirely optional, providing users with an opt-in way for custom kubectl behaviour without impacting any existing setups.

Kubernetes 1.34 – What you need to know

Large Language Models (LLMs) are now at the cutting edge of mainstream AI systems. Their impact has been seismic, sparking a new gold rush as application developers transform the user experience away from clicks and commands into natural language and advanced automation

However, application developers have a barrier to overcome. AI models need data to reason and respond to a particular application domain. Say, for example, the user of an artifact management platform needs to pose a simple question, ‘What packages are trending upward in usage?’ To answer this question, the AI needs to be trained on package data, including usage data.

In general, an AI model's capabilities depend on the information, or context, on which it is trained. And here is the problem - AI models tend to be isolated from data. Getting data into the AI model is not trivial. The data is often behind information silos and legacy systems. Integrating models usually involves a custom implementation with bespoke connectors.

It’s as if each time you plugged a device into a computer’s USB port, you had to configure the connection on a per-device basis. And so it is with AI models. The implementation has to be done on a per-data source, per-AI model basis, each time needing its own connector, protocol, and security checks. This multiplies the work involved in integration and tethers the LLM to a particular source, making it difficult to scale.

This multiplicity of connectors between each model and each source has been termed the MxN problem. Every model (M) needs a custom integration for every source or service (N), creating a multitude of one-off adapters that create a build and maintenance headache. As more models and tools enter the ecosystem, the integration load grows exponentially.

Luckily, Anthropic devised a solution to solve the mess of custom integrations that results from the MxN approach. The Model Context Protocol (MCP) is an open standard that provides a universal way of connecting AI models to external data sources and systems.

MCP provides a common interface, regardless of data source, eliminating the need for custom integrations for models and sources. The use of MCP as a middle layer reduces the MxN problem to a much more scalable M+N, as each model and source only has to integrate with MCP.


The USB-C port analogy is often used for MCP, as it, too, uses a single connectivity standard to provide a universal way to connect different devices. In fact, any common protocol has the same kind of utility. HTTP created a universal language for web browsers and servers to interact. MCP enables interoperability across platforms like OpenAI, Anthropic, Google DeepMind, and Microsoft, giving AI assistants access to tools and repositories in a secure and standardized way.

MCP is built using a client-server architecture, and works as follows:

The AI model (host) uses connectors (clients) to send requests for information or actions using MCP to an MCP server.

The server receives the requests, acting as an abstraction layer in front of the data source, exposing a controlled set of functions.

The server then translates the model’s requests into safe, scoped operations. This means data is accessed securely and consistently, and the model doesn’t need to know server implementation details for each data source.

The server capabilities are exposed via JSON-ROC methods, which are mapped to one or more internal database queries or functions. These could include file lookups, data queries, transformation functions, or any other tool-like operation.

The AI model can introspect the available methods and call them with specific parameters. The server authenticates the request and calls the appropriate backend query using its own connection to the database. A structured response is then returned to the model. This completely decouples the database logic from the AI model.

MCP provides powerful capabilities through code execution and data access. To address this, MPC includes a series of recommendations for best practices for securing data within your infrastructure.

These best practices are shaped around four key principles:

- Users must explicitly consent to and understand what data is accessed and what actions are performed. Implementations should offer clear UI elements to authorize and review activities.

- Resource data must not be transmitted or exposed without user approval. All access must be explicitly granted and protected through proper controls.

 - Since tools may trigger arbitrary code execution, hosts must treat them with caution and verify trust. Descriptions alone should not be trusted unless verified, and execution must be explicitly authorized.

 - Any use of LLM sampling must be user-approved. Users should be able to control whether sampling occurs, what prompt is sent, and what response visibility the server has.

MCP is unable to enforce these security principles; that is the responsibility of developers. And as a rule, they should follow these guidelines as outlined in the MCP spec:

The application should use robust flows for consent and authorization

All security implications must clearly be outlined in documentation

Appropriate access controls and data protection must be in place

Follow all security best practices when building integrations

When designing features, always consider privacy implications

Some diligence is required to ensure appropriate security controls are in place.

Using AWS CodeBuild to push packages to Cloudsmith

More articles

Python 3.14 – What you need to know

Migrating from Docker Content Trust to Sigstore

Kubernetes 1.34 – What you need to know

Adding AI to applications using the Model Context Protocol

Docker Hardened Images & Cloudsmith: Modern Security for the Software Supply Chain

Secure Docker Image Pulls from Cloudsmith to Kubernetes using OIDC