<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Daniel Adeboye on Medium]]></title>
        <description><![CDATA[Stories by Daniel Adeboye on Medium]]></description>
        <link>https://medium.com/@adeboyedn?source=rss-751549ff0fd3------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*UjWnIwqkl88bcj3mtNYCRA.jpeg</url>
            <title>Stories by Daniel Adeboye on Medium</title>
            <link>https://medium.com/@adeboyedn?source=rss-751549ff0fd3------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 08:43:34 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@adeboyedn/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Agent framework vs agent harness: What’s the difference?]]></title>
            <link>https://medium.com/@adeboyedn/agent-framework-vs-agent-harness-whats-the-difference-3ce071ce3ee8?source=rss-751549ff0fd3------2</link>
            <guid isPermaLink="false">https://medium.com/p/3ce071ce3ee8</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[agent-harness]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[agent-framework]]></category>
            <dc:creator><![CDATA[Daniel Adeboye]]></dc:creator>
            <pubDate>Fri, 20 Nov 2026 09:35:32 GMT</pubDate>
            <atom:updated>2026-11-20T09:36:54.551Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LdXw__rtlG9GOwYicOhgbw.png" /></figure><p>Building an AI agent and running one are two different problems. That’s where agent frameworks and agent harnesses come in.</p><p>Frameworks like LangChain, LangGraph, and CrewAI give you the building blocks to create agents. Tools like Claude Code, Codex, and OpenCode come with harnesses that manage how agents execute tasks.</p><p>The terms are often used interchangeably, but they serve different purposes. And understanding the difference raises another important question: where does your agent actually run?</p><h3>What is an agent framework?</h3><p>An agent framework is a set of building blocks for creating your own AI agents. It provides components such as model integrations, tool definitions, memory, and workflow orchestration, which you can combine to build agents for different tasks.</p><p>LangChain describes itself as a framework that provides the core building blocks for developing agents. LangGraph and CrewAI are other examples, each with its own approach to defining workflows and coordinating agents.</p><p>With a framework, you have control over how your agent works. You decide when the model gets called, which tools it can use, what happens when a tool fails, and when a task is complete.</p><p>That flexibility is useful when you’re building custom agent applications. But it also means you’re responsible for putting the pieces together and maintaining the logic.</p><p>Think of it like a box of LEGO bricks. You can build almost anything, but you have to decide what to build and put the pieces together yourself.</p><h3>What is an agent harness?</h3><p>An agent harness is the software that surrounds a language model and manages its execution as an agent. It handles the loop: send context to the model, execute the tool it requests, feed the result back, and repeat until the task is complete.</p><p>A harness can also handle other important parts of agent execution, such as managing context, requesting approval before risky actions, handling errors, and controlling which tools and resources the agent can access.</p><p>Claude Code, Codex, and OpenCode are examples of coding agents that come with their own harnesses. You don’t have to build the agent execution loop yourself. You give the agent a task, and it manages the process of working through it.</p><p>It’s worth distinguishing the harness software from the environment where it runs. The harness manages the agent’s execution, but it still needs somewhere to run, access to the files and tools it needs, and appropriate permissions.</p><p>A simple way to remember the difference: a framework gives you the building blocks to create an agent, while a harness provides the software that runs and manages one.</p><h3>Agent framework vs agent harness: What’s the difference?</h3><p>Agent frameworks and harnesses approach agent development from different angles. A framework gives you the building blocks to create an agent, while a harness provides the software that manages its execution.</p><p>Here’s how they compare:</p><p><strong>Primary purpose</strong></p><p>An agent framework helps you build and structure agents. It provides components for model integrations, tools, memory, and workflow orchestration.</p><p>An agent harness manages agent execution. It handles the execution loop, context management, tool calls, and error recovery.</p><p><strong>What developers do</strong></p><p>With a framework, you define the agent’s logic and assemble its components. You decide which tools it can use, how its workflow is structured, and what should happen at each step.</p><p>With a harness, much of the execution logic is already provided. You configure the agent, give it instructions, and control its access to tools.</p><p><strong>Examples</strong></p><p>Common agent framework examples include LangChain, LangGraph, and CrewAI. Examples of coding agents that come with their own harnesses include Claude Code, Codex, and OpenCode.</p><p>The key difference is how much of the agent’s execution you need to build yourself. With a framework, you put the components together and define how the agent works. With a harness, the execution loop is already handled, so you can focus on the task and how the agent should behave.</p><p>Neither approach is inherently better. It depends on what you’re trying to build.</p><h3>Can you use an agent framework and harness together?</h3><p>Yes. In fact, some harnesses are built on top of existing agent frameworks.</p><p>Deep Agents is a good example. LangChain describes it as an agent harness built on top of LangChain and LangGraph. It adds capabilities such as planning, a virtual filesystem, subagents, and long-term memory, so developers don’t have to implement these pieces themselves.</p><p>Microsoft takes a similar approach with its Agent Framework, which includes a Harness built from the framework’s existing components. Rather than introducing an entirely separate runtime, it extends the framework’s capabilities for running agents.</p><p>Anthropic takes a different approach. The <a href="https://code.claude.com/docs/en/agent-sdk/overview">Claude Agent SDK</a> exposes the agent capabilities behind Claude Code, allowing developers to build applications that use the same underlying execution loop with Python or TypeScript.</p><p>These examples show that frameworks and harnesses don’t have to be separate products. A framework can provide the components used to build a harness, while a harness can package those components into a more complete agent execution system.</p><p>The distinction is less about which tool you choose and more about what it provides: building blocks for creating agents, or an execution system that manages how they work.</p><h3>Common misconceptions about agent frameworks and harnesses</h3><ul><li><strong>Harness is just a new name for framework:</strong> It isn’t. A framework helps you build an agent. A harness is an agent loop that already runs. Some products are both, but the jobs are different.</li><li><strong>Harnesses are less flexible:</strong> Many harnesses let you customise their behaviour by adding tools, changing prompts, and configuring permissions. The level of customisation depends on the harness.</li><li><strong>You have to pick one:</strong> As we just saw, plenty of harnesses are built on frameworks. You can start with a harness and drop down to the framework when you need more control.</li><li><strong>The harness keeps the agent safe:</strong> This is the big one. A harness can request approval before running a risky command, but approval alone doesn’t provide isolation. Once a command executes, its access depends on the permissions and resources available in the underlying environment. The Deep Agents README says it plainly: enforce boundaries at the tool or sandbox level, instead of expecting the model to police itself. Which brings us to a question most comparisons skip.</li></ul><h3>When should you use an agent framework vs an agent harness?</h3><p><strong>Use an agent framework if:</strong></p><ul><li>You’re building an agent into your own product or application.</li><li>You need control over the agent’s logic, tools, and execution flow.</li><li>Your workflow requires custom orchestration that existing harnesses don’t support.</li></ul><p><strong>Use an agent harness if:</strong></p><ul><li>You want to get an agent working on a task without building the execution loop yourself.</li><li>You’re running coding, research, or other multi-step tasks.</li><li>You’d rather configure an existing agent than build one from scratch.</li></ul><p>If you’re somewhere in between, you can start with a programmable option such as Deep Agents or the Claude Agent SDK. You get existing agent execution capabilities while retaining the ability to customise the tools, workflows, and behaviour for your use case.</p><h3>Where should your agent run?</h3><p>Here’s the part most comparisons between agent frameworks and harnesses skip.</p><p>The harness manages how an agent executes tasks, but it doesn’t automatically determine what commands it can run or access. That depends on the environment where the agent is running and the permissions it has.</p><p>Run Claude Code on your laptop, for example, and its commands may have access to your local files, SSH keys, and credentials, depending on how you’ve configured its permissions. Run an agent built with a framework inside your application server, and the code it executes may have access to resources on that server.</p><p>So whether you’re using a framework, a harness, or both, there’s another question to answer: where should your agent run?</p><p><a href="https://northflank.com/">Northflank</a> is the runtime platform for autonomous software. It provides the infrastructure to run AI agents, the code they generate, and the applications they build, from isolated sandboxes to production workloads.</p><p>For coding agents, Northflank provides <a href="https://northflank.com/product/cloud-agents"><strong>Cloud Harnesses</strong></a>, managed cloud workspaces where you can run agents such as Claude Code, Codex, OpenCode, Pi, Cursor, or bring your own agent.</p><p>Unlike the agent harnesses discussed earlier, <a href="https://northflank.com/product/cloud-agents">Northflank Cloud Harnesses</a> don’t manage the agent’s execution loop. Instead, they provide the runtime environment where the agent operates.</p><p>Here’s what you get:</p><ul><li><strong>Isolated execution:</strong> Each Cloud Harness runs in its own Kata microVM, separated from other workloads.</li><li><strong>Persistent workspaces:</strong> Files and installed dependencies remain available across pauses and restarts when workspace persistence is enabled.</li><li><strong>Multiple coding agents:</strong> Run supported agents or start with an empty harness and configure your own tools.</li><li><strong>Pause and resume:</strong> Pause a harness to stop compute usage, then resume it later with your workspace files intact. Running processes stop when the harness is paused.</li><li><strong>Bring your own cloud:</strong> Run harnesses on Northflank’s cloud or inside your own infrastructure using BYOC.</li></ul><p>You can learn more in the <a href="https://northflank.com/docs/v1/application/cloud-harness/quickstart">Northflank Cloud Harness documentation</a>.</p><p>For developers building their own agents, Northflank also provides <a href="https://northflank.com/product/sandboxes">Sandboxes</a>. These let agents create isolated microVM environments through an API to execute generated code without running it directly on the application server.</p><p>The idea is to let your agent handle the task while providing the commands it runs with an isolated environment and the resources and permissions they need.</p><h3>Conclusion</h3><p>Agent frameworks and harnesses solve different problems, but they can work together.</p><p>A framework gives you the building blocks to create an agent. A harness manages how that agent executes tasks. And the environment where it runs determines what resources and systems it can access.</p><p>Choose the tools that give you the right balance of control and flexibility, but don’t overlook the execution environment. Building an agent that works is one thing. Making sure it runs in an environment with the right isolation, permissions, and resources is another.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3ce071ce3ee8" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How much does it cost to run AI coding agents in the cloud?]]></title>
            <link>https://medium.com/@adeboyedn/how-much-does-it-cost-to-run-ai-coding-agents-in-the-cloud-2f6ff7cf9ec9?source=rss-751549ff0fd3------2</link>
            <guid isPermaLink="false">https://medium.com/p/2f6ff7cf9ec9</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[Daniel Adeboye]]></dc:creator>
            <pubDate>Mon, 26 Oct 2026 18:31:39 GMT</pubDate>
            <atom:updated>2026-11-20T09:38:49.735Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*B75ZrgiibO0_CdZuOrPKkw.png" /></figure><p>Coding agents such as Claude Code and Codex increasingly run in the cloud rather than on developers&#39; laptops. They can keep working after you close your laptop, run in isolated environments, and handle multiple tasks in parallel. But every environment has a cost, and the actual bill depends on more than the advertised hourly rate.</p><p>This guide breaks down the cost of running AI coding agents in the cloud, from infrastructure and model usage to monthly costs across different workloads. It also compares cloud sandbox providers, explains what drives costs up, and covers how to keep spending under control. Pricing figures are based on <a href="https://github.com/northflank-examples/agent-sandbox-pricing">publicly available provider pricing</a>, with estimates calculated using clearly stated workload assumptions.</p><h3>What makes up the cost of an AI coding agent?</h3><ol><li><strong>Model usage:</strong> The agent calls a language model at every step. This is billed by the model provider, either through a subscription plan with usage limits or per token through an API key, and it is the same wherever the agent runs.</li><li><strong>Compute:</strong> CPU and memory for the environment the agent runs in. Most platforms bill the CPU and memory allocated to the environment for every second it runs, whether the agent is busy or waiting on a model response.</li><li><strong>Storage:</strong> The persistent disk that keeps the repository, installed packages, and work in progress between sessions. Ephemeral storage is often included in the compute price.</li><li><strong>Network egress</strong>: Data sent out of the environment, such as pushing code or build artifacts to external services or sending requests to external APIs.</li><li><strong>Plan fees and concurrency limits:</strong> Some platforms charge a monthly plan fee or require a higher plan to run more environments simultaneously.</li><li><strong>Idle time:</strong> An environment left running after the agent finishes keeps billing for compute until it is paused or deleted. For long-lived agent environments, idle time is often the largest part of the bill.</li></ol><h3>How much do coding agent models cost?</h3><p>Model costs are billed by the model provider, not the cloud platform. Coding agents use one of two pricing models.</p><ul><li><strong>Subscription plans:</strong> include agent usage subject to each plan’s limits. Claude Code is included with Claude subscription plans, and Codex is included with eligible ChatGPT plans. A subscription can suit individual developers with steady daily usage.</li><li><strong>API keys:</strong> bill per input and output token at the provider’s API rates. API billing suits automated workloads, teams that want usage-based costs, and agents routed through cloud providers such as Amazon Bedrock, Google Cloud, or Microsoft Foundry, where usage is billed through the cloud account.</li></ul><p>Running an agent in the cloud does not inherently change its model costs or usage limits. The same account and applicable limits generally apply whether you run it on a laptop or in a cloud environment.</p><h3>How much does the cloud environment cost?</h3><p>The table below compares the cost of running a 2 vCPU, 4 GiB environment across cloud sandbox platforms, assuming CPU utilization averages 25%.</p><p><strong>Northflank</strong></p><ul><li>CPU: $0.01667 per vCPU-hour</li><li>Memory: $0.00833 per GiB-hour</li><li>Estimated cost: $0.0667/hour</li><li>Billing: Allocated resources</li></ul><p><strong>Cloudflare Sandboxes</strong></p><ul><li>CPU: $0.072 per vCPU-hour</li><li>Memory: $0.009 per GiB-hour</li><li>Estimated cost: $0.0900/hour</li><li>Billing: Active CPU</li></ul><p><strong>Sprites (Fly.io)</strong></p><ul><li>CPU: $0.0385 per vCPU-hour</li><li>Memory: $0.02187 per GiB-hour</li><li>Estimated cost: $0.1067/hour</li><li>Billing: Usage-based</li></ul><p><strong>Vercel Sandbox</strong></p><ul><li>CPU: $0.128 per vCPU-hour</li><li>Memory: $0.0212 per GiB-hour</li><li>Estimated cost: $0.1488/hour</li><li>Billing: Active CPU</li></ul><p><strong>E2B</strong></p><ul><li>CPU: $0.0504 per vCPU-hour</li><li>Memory: $0.0162 per GiB-hour</li><li>Estimated cost: $0.1656/hour</li><li>Billing: Allocated resources</li></ul><p><strong>Daytona</strong></p><ul><li>CPU: $0.0504 per vCPU-hour</li><li>Memory: $0.0162 per GiB-hour</li><li>Estimated cost: $0.1656/hour</li><li>Billing: Allocated resources</li></ul><p><strong>Modal</strong></p><ul><li>CPU: $0.07096 per vCPU-hour</li><li>Memory: $0.02401 per GiB-hour</li><li>Estimated cost: $0.2380/hour</li><li>Billing: Maximum of requested and actual usage</li></ul><p>Prices verified in October 2026 from each vendor’s pricing page. See the <a href="https://github.com/northflank-examples/agent-sandbox-pricing">full comparison</a> for sources, storage, egress, and GPU pricing.</p><p>The billing model changes what you pay for. <strong>Allocated</strong> billing charges for the CPU and memory reserved for the environment every second it runs, busy or not. <strong>Active CPU</strong> billing charges memory while the environment runs but CPU only while it executes, which suits agents that spend most of their time waiting on model responses but costs more for busy workloads such as builds and tests. <strong>Usage</strong> billing charges only for the CPU time and memory actually consumed.</p><h3>What does it cost in practice?</h3><p>These examples use <a href="https://northflank.com/">Northflank’s</a> list prices for a 2 vCPU, 4 GiB Harness at $0.0667 per hour and a persistent disk at $0.11657 per GiB per month. Model usage is excluded and billed separately by the model provider.</p><p><strong>1. One developer running one agent</strong></p><ul><li>Usage: 4 hours/day, 22 days/month</li><li>Storage: 10 GiB</li><li>Estimated monthly cost: <strong>$7</strong></li></ul><p><strong>2. Five developers running two agents each</strong></p><ul><li>Usage: 10 agents, 4 hours/day, 22 days/month</li><li>Storage: 10 GiB per environment</li><li>Estimated monthly cost: <strong>$70</strong></li></ul><p><strong>3. Ten always-on environments</strong></p><ul><li>Usage: 730 hours/month per environment</li><li>Storage: 20 GiB per environment</li><li>Estimated monthly cost: <strong>$510</strong></li></ul><p><strong>4. Ten environments paused when idle</strong></p><ul><li>Usage: 8 hours/day, 22 days/month</li><li>Storage: 20 GiB per environment</li><li>Estimated monthly cost: <strong>$141</strong></li></ul><p><strong>5. One hundred short-lived tasks</strong></p><ul><li>Usage: 100 tasks, 1 hour each</li><li>Storage: None</li><li>Estimated monthly cost: <strong>$7</strong></li></ul><p>The difference between the always-on and paused rows is the cost of idle time. The environments do the same work in both cases, but leaving them running outside working hours costs about $369 more per month. The same pattern holds on every platform that bills while an environment runs.</p><p>To price your own workload across platforms, use the <a href="https://p01--sandbox-pricing-estimator--wdgpv6bn4jz8.code.run/">agent sandbox pricing estimator</a>.</p><h3>What drives the cost of coding agents up, and how do you reduce it?</h3><p><strong>Idle environments</strong></p><p>Cloud environments can continue costing money even when an agent isn’t doing anything, depending on the billing model. Pause environments between tasks or configure them to shut down automatically when they’re no longer needed.</p><p><strong>Oversized environments</strong></p><p>Giving every agent more CPU and memory than it needs increases infrastructure costs. Start with smaller configurations and scale up when workloads require more resources.</p><p><strong>Running multiple agents in parallel</strong></p><p>More concurrent agents mean more environments, resource consumption, and potentially higher plan fees. Set concurrency limits and run additional agents only when the extra parallelism is worth the cost.</p><p><strong>Persistent storage</strong></p><p>Workspaces, repositories, dependencies, and build artifacts can accumulate storage costs over time. Set appropriate storage limits and remove data that no longer needs to be retained.</p><p><strong>Long-running sessions</strong></p><p>Agents can keep environments active long after a task has finished. Use session timeouts, automate cleanup, and avoid keeping environments alive without a clear reason.</p><p><strong>Network egress</strong></p><p>Sending code, build artifacts, and other data out of an environment can add transfer costs. Keep related services close together where possible and avoid unnecessary data transfers.</p><p><strong>Plan fees and concurrency limits</strong></p><p>Some providers charge monthly plan fees or restrict how many environments you can run simultaneously. Include these costs when comparing providers, especially if you’re running agents across a team.</p><p>The biggest savings usually come from matching resource allocation to actual workload, pausing environments when idle, and choosing a billing model that fits how your agents run.</p><h3>How do you choose a cloud platform for AI coding agents based on cost?</h3><ul><li><strong>Billing model:</strong> Allocated billing can be cost-effective for agents that run builds and tests continuously. Active CPU or usage-based billing may suit workloads that spend more time idle.</li><li><strong>Idle behaviour:</strong> Check what a platform charges when an environment is idle and whether you can pause it. Pausing environments to stop compute charges can significantly reduce costs for long-running agents.</li><li><strong>Plan fees and concurrency:</strong> If you run multiple agents simultaneously, compare the cost of the plan needed to support that concurrency, not just the hourly rate.</li><li><strong>Session limits:</strong> Agents that run for hours need a platform that supports long sessions, or you may need additional logic to restart them and restore their state.</li><li><strong>Storage and egress:</strong> Persistent disks, snapshots, and network egress can add up for environments that run for weeks.</li><li><strong>BYOC pricing:</strong> Running in your own cloud account lets you use your cloud provider’s rates and any applicable committed-spend discounts. Check the platform’s BYOC management fees as well as your underlying cloud bill.</li><li><strong>Isolation:</strong> MicroVM isolation provides a stronger boundary for environments running model-generated commands. Check what isolation each platform includes before comparing prices.</li></ul><p>For a broader comparison of features and providers, see <a href="https://northflank.com/blog/best-cloud-environments-for-ai-coding-agents">Best cloud environments for AI coding agents</a>.</p><h3>Running AI coding agents on Northflank</h3><p><a href="https://northflank.com/product/cloud-agents">Northflank</a> takes a straightforward approach to running coding agents in the cloud. Its Cloud Harnesses provide dedicated environments for Claude Code, Codex, OpenCode, Pi, Cursor, or your own agent. Each Harness runs in its own Kata microVM, with a persistent workspace, repository access, and the ability to keep working after you disconnect from your machine.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JS8_irBsj04EcR7ox2nrBA.png" /></figure><p>The <a href="https://northflank.com/pricing">pricing model</a> is particularly relevant for workloads that run multiple agents in parallel. Northflank bills by the second, with no seat or per-agent fees. On managed cloud, CPU costs $0.01667 per vCPU-hour, and memory costs $0.00833 per GB-hour. A 2 vCPU, 4 GB environment therefore costs about $0.0667 per hour, excluding persistent storage and any applicable additional charges.</p><p>You can also pause a Harness when it is not needed. Compute billing stops while its workspace remains available on persistent storage, so you don’t have to keep paying for an idle environment just to preserve your work. The free Sandbox plan includes a Harness with up to 1 vCPU and 1 GB of memory.</p><p>Another option is <a href="https://northflank.com/product/bring-your-own-cloud">bring-your-own-cloud (BYOC)</a>. You can run Harnesses in your own AWS, Google Cloud, Azure, Oracle, or CoreWeave environment. Your cloud provider bills the underlying infrastructure, while <a href="https://northflank.com/pricing">Northflank charges</a> a management fee of $0.02083 per node vCPU-hour ($15.00/ node vCPU / month). Because this fee is based on managed node capacity, adding more Harnesses to existing nodes does not directly increase the management fee.</p><p>The bigger advantage is that coding agents aren’t isolated from the rest of your development infrastructure. Through the Northflank <a href="https://northflank.com/docs/v1/application/ai-agents/agent-skills">Skills</a> and <a href="https://northflank.com/docs/v1/application/ai-agents/mcp-server">MCP</a>, agents can create sandboxes, deploy services, provision databases, and generate preview environments for the code they write. That means you can run agents and the infrastructure they need on the same platform, rather than managing separate services for coding, testing, and deployment. You can explore the <a href="https://northflank.com/product/cloud-agents">Northflank Cloud Agents</a> page for details.</p><h3>Conclusion</h3><p>The biggest factor in your cloud coding agent bill isn’t always the hourly rate. It’s how long your environments run, how much capacity they reserve, and whether you pay for them while they’re idle.</p><p>If you’re running agents for a few hours a day, pausing environments between tasks can make a significant difference. If you’re running dozens of agents in parallel, per-agent fees, concurrency limits, and storage costs become just as important as the compute rate.</p><p>The goal is to choose a platform that fits how your agents actually work, rather than paying for infrastructure they don’t need.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2f6ff7cf9ec9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Startups Should Care About Security From Day Zero.]]></title>
            <link>https://medium.com/@adeboyedn/why-startups-should-care-about-security-from-day-zero-a1ad169cd4ed?source=rss-751549ff0fd3------2</link>
            <guid isPermaLink="false">https://medium.com/p/a1ad169cd4ed</guid>
            <dc:creator><![CDATA[Daniel Adeboye]]></dc:creator>
            <pubDate>Thu, 08 Feb 2024 12:35:17 GMT</pubDate>
            <atom:updated>2024-02-08T12:35:17.023Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iOmIxe641JgC8YORbA52sg.png" /></figure><p>Whispers in my ear: <strong><em>“These founders do not prioritize security until there is a reported security breach.”</em></strong> This sentiment has persisted for several decades. I said to myself, <strong><em>“It’s not that they don’t care; they just don’t feel the urgency, especially in their early stages.”</em></strong> Whether you are in your early stages or not, it’s crucial to prioritize security from the outset to maintain user trust and mitigate future risks.</p><p>Reader’s thoughts: <strong><em>“Why discuss this now? Are you attempting to promote </em></strong><a href="https://gitsecure.dev/"><strong><em>Gitsecure</em></strong></a><strong><em>, your developer security startup?”</em></strong></p><p>My first response: “<strong><em>Why discuss this now?”</em></strong> It’s because I’ve observed a concerning trend among many startups I admire. They often announce security breaches, which instill fear in their users and cost them a lot of money. One recent incident involved Resend, a company I greatly respect. Let’s talk about what happened to Resend, what lessons can be learned, and how such breaches could have been prevented from the start. Furthermore, we’ll explore the measures they’ve taken to prevent future breaches.</p><h3>What Happened with Resend?</h3><h3>Resend on Twitter: &quot;On January 7th, attackers used a leaked environment variable of the Resend database API to access customer data.Here&#39;s a summary of what went wrong, how the incident was resolved, and the work we are doing to ensure it does not happen again.https://t.co/8cFCk0nc8E / Twitter&quot;</h3><p>On January 7th, attackers used a leaked environment variable of the Resend database API to access customer data.Here&#39;s a summary of what went wrong, how the incident was resolved, and the work we are doing to ensure it does not happen again.https://t.co/8cFCk0nc8E</p><p>Above is a tweet from the Resend team regarding an incident where attackers exploited a leaked environment variable to access customer data, which includes emails sent, domains, API keys (encrypted), logs, and contacts, affecting all users. They also stated that the actual content of the emails was not accessed, nor were any unencrypted private keys for the Resend API or DKIM. You can learn more <a href="https://resend.com/blog/incident-report-for-january-10-2024">here</a>.</p><h3>Lessons Learned:</h3><ol><li>Avoid exposing critical information, like API keys, on the client side. Employ secure practices, such as encryption, secure vaults, or key management systems, to store and retrieve such information.</li><li>Promptly investigate and address unusual patterns in system behavior, even during non-standard working hours. Thoroughly examine any deviations from normal operations.</li><li>Implement robust monitoring and alerting systems to detect unusual activities in real time. Early detection is crucial for minimizing the impact of a security breach.</li><li>Communicate openly and transparently with users about the nature and impact of the breach. Providing detailed incident reports builds trust and helps users understand the steps taken for remediation.</li><li>Take immediate and proactive actions in response to a security incident, including removing exposed credentials, rotating access keys, and enforcing Multi-Factor Authentication (MFA).</li></ol><h3>How Could Resend Have Prevented the Incident?</h3><ol><li>Resend could have prevented the incident if they had monitoring and alerting systems in place for API access, which would promptly notify the security team of any suspicious activities.</li><li>Resend could have prevented the incident if they had employed secure practices, such as encryption, using secure vaults, or key management systems, to store and retrieve their secret credentials.</li><li>Resend could have prevented the incident by embracing the principle of least privilege through the enforcement of strict access controls. It is essential to consistently review and update access permissions to minimize potential risks.</li><li>Resend could have prevented the incident if they had regularly rotated access credentials and API keys from time to time, even without an indication of a breach, to minimize the risk of unauthorized access.</li><li>Resend could have prevented the incident if they had enforced MFA across all systems touching the database, adding an extra layer of security against unauthorized access.</li><li>Resend could have prevented the incident if they had partnered with third-party cybersecurity companies for regular security assessments, ensuring resilience to emerging threats.</li></ol><h3>What has the Resend team done so far to prevent future breaches?</h3><p>The Resend team has taken action on the security breach and has partnered with a security provider that will help them prevent security incidents like that in the future, as stated in their blog post. Read more <a href="https://resend.com/blog/incident-report-for-january-10-2024">here</a>, They have also made everything around security at Resend transparent for their users to see, which will helps build trust.</p><iframe src="https://cdn.embedly.com/widgets/media.html?type=text%2Fhtml&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;schema=twitter&amp;url=https%3A//twitter.com/resend/status/1753079565390590165%3Fs%3D20&amp;image=" width="500" height="281" frameborder="0" scrolling="no"><a href="https://medium.com/media/4b13ba181c3a310292c3b5a479db4bd2/href">https://medium.com/media/4b13ba181c3a310292c3b5a479db4bd2/href</a></iframe><p>Along the line, Zeno, who is the founder and CEO of Resend tweeted, asking about security tools and vendors. Of course, I had to step in and talk about <a href="https://gitsecure.dev/">Gitsecure</a>.</p><h3>Daniel Adeboye on Twitter: &quot;I recommend checking out @gitsecure . It can assist your team in developing secure applications by offering code vulnerability detection, identifying leaked secret credentials, and providing recommendations for fixing any vulnerabilities it detects in your codebase. / Twitter&quot;</h3><p>I recommend checking out @gitsecure . It can assist your team in developing secure applications by offering code vulnerability detection, identifying leaked secret credentials, and providing recommendations for fixing any vulnerabilities it detects in your codebase.</p><p>My second response: <strong><em>Are you attempting to promote </em></strong><a href="https://gitsecure.dev/"><strong><em>Gitsecure</em></strong></a><strong><em>?</em></strong></p><p>Of course, I’m trying to promote the value that Gitsecure can offer. With Gitsecure in your workflow, every startup is assured that they are shipping quality and secure codes.</p><h3>What exactly is Gitsecure?</h3><p><a href="https://gitsecure.dev/">Gitsecure</a> is an AI-powered developer security platform for modern teams. Our mission is to assist you in creating and delivering high-quality and secure code. Below are some of the features that Gitsecure provides:</p><ol><li>Static Application Security Testing (SAST)</li><li>Secrets Detection</li></ol><p>Watch the 30-second GIF below to see how Gitsecure works.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*EqbPM0cBgxMGrO1CKaIbeA.gif" /></figure><h3>Conclusion</h3><p>In conclusion, the Resend incident serves as a reminder that security cannot be an afterthought for startups. Prioritizing security from day <strong><em>zero </em></strong>is not just a precaution but a strategic imperative to safeguard user trust and the longevity of the business. Remember, secure coding isn’t just a necessity; it’s a competitive advantage.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a1ad169cd4ed" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>