AI coding tools risks became painfully concrete when a developer reported that Claude, an AI coding assistant, deleted over 48,000 files in just 103 seconds. For small and mid-sized business owners watching their teams adopt AI development tools, this incident raises an urgent question: do you know what your developers’ AI assistants can actually do to your codebase, your data, and your business?
The answer matters because AI coding assistants now operate with the same system permissions as the humans who invoke them. They can read your entire repository, modify production code, delete files, and push changes to live systems. All in seconds. All without a second approval. And all with the best of intentions, because these tools are doing exactly what they were designed to do: automate tasks at machine speed.
What happened when an AI coding assistant deleted 48,000 files?
A developer using Claude’s code agent asked it to perform what seemed like a routine task. Within 103 seconds, the AI had deleted 48,000 files. The developer posted about the incident publicly, and the story spread through the software development community like wildfire.
The details matter. This was not a malicious attack. The AI was not hacked. It was working as designed, interpreting instructions and executing them with ruthless efficiency. The problem was that the instruction, as understood by the AI, had consequences the developer did not anticipate or want.
For an SMB owner, the lesson is this: AI coding tools risks include the speed and scale at which mistakes happen. A human developer making the same error might delete a few files before noticing something was wrong. An AI can delete tens of thousands before you can reach for the keyboard.
Recovery from that kind of incident is not trivial. If you have good backups and version control, you might restore the files in a few hours. If you do not, or if the AI modified files instead of deleting them, you could spend days or weeks sorting out what changed, what broke, and what data was lost. At $150 per hour for developer time, a 20-hour recovery effort costs $3,000. A 40-hour effort costs $6,000. And that assumes you catch it immediately.
Do AI code assistants create security exposure for your business?
Yes, and the exposure is broader than most business owners realize. AI coding assistants can introduce at least four categories of risk to your operation.
First, there is the permission problem. When your developer grants an AI assistant access to your code repository, the AI inherits whatever permissions that developer has. If your developer can push code to production, so can the AI. If your developer can access customer databases for debugging, so can the AI. Most AI tools do not have separate permission models or approval gates. They act as extensions of the user.
Second, there is the data leakage problem. Some AI coding tools send code snippets to external servers for processing. If your repository contains API keys, database credentials, customer data, or proprietary algorithms, those details might leave your network. Depending on the tool and its configuration, you may or may not have control over where that data goes, how long it is retained, or who else can see it.
Third, there is the code quality problem. AI-generated code can contain security vulnerabilities, outdated libraries, or logic errors that your team might not catch in review. If that code makes it to production, you could face data breaches, downtime, or compliance violations. For businesses in regulated industries (healthcare under HIPAA, financial services under the Gramm-Leach-Bliley Act, or defense contractors under CMMC), a single AI-introduced flaw could trigger an audit failure or regulatory penalty.
Fourth, there is the accountability problem. When something goes wrong, who is responsible? The developer who invoked the AI? The vendor who built the tool? Your IT leadership for allowing it? The answer affects your insurance coverage, your legal exposure, and your ability to learn from the incident.
What controls should SMBs put in place before developers use AI coding tools?
You need a written policy, technical guardrails, and a clear approval process. Here is what that looks like in practice.
Start with a written acceptable-use policy that names the AI tools your business has approved for development work. This is not about banning AI. It is about making a conscious choice. List the specific tools (GitHub Copilot, ChatGPT, Claude, Cursor, or others), specify what tasks they can be used for (code suggestions, documentation, test generation), and define what is off-limits (production deployments without human review, processing of live customer data, or access to credentials).
Next, implement technical controls. Configure your version control system to require human review before any code reaches production. Use branch protection rules to prevent direct commits to main branches. Set up automated scanning to catch secrets (API keys, passwords) before they are committed. Restrict AI tools to test environments or sandboxes where mistakes cannot affect live systems.
Third, establish a data boundary. If your business handles regulated data (health records, financial information, personal data under GDPR or CCPA), configure your AI tools to operate in offline or self-hosted modes, or prohibit their use on repositories containing sensitive data. Some vendors offer enterprise versions with data residency guarantees and no-training clauses. Evaluate whether that investment is necessary for your risk profile.
Fourth, train your team. Make sure every developer understands what the AI can and cannot do, what the business rules are, and how to recognize when an AI suggestion introduces risk. A 30-minute training session costs far less than a single data breach or compliance incident.
Finally, audit usage. Review logs periodically to confirm that AI tools are being used in approved ways. If you discover that a developer is using an unapproved tool or bypassing guardrails, address it immediately. Policies without enforcement are just suggestions.
How much does an AI coding mistake actually cost a small business?
The direct costs are developer time, recovery effort, and potential data loss. The indirect costs are harder to measure but often larger.
If an AI assistant deletes critical files, you will spend 10 to 40 hours recovering them, testing the restored system, and verifying nothing else broke. At $100 to $200 per hour for developer time, that is $1,000 to $8,000 in labor.
If the AI introduces a security vulnerability that leads to a data breach, the costs escalate quickly. The average cost of a small business data breach in 2024 was $120,000, according to IBM’s Cost of a Data Breach Report. That includes forensics, notification, legal fees, and lost business. For an SMB with tight margins, a breach of that size can be existential.
If the mistake affects a client project, you may face contractual penalties, lost revenue, or reputational damage. Clients who learn that your team lost control of an automated tool may question your overall security posture. Prospective clients who hear about the incident may choose a competitor.
The less obvious cost is opportunity cost. Every hour your team spends recovering from an AI mistake is an hour not spent building new features, serving customers, or growing the business. For a five-person development team, losing a week to recovery means losing 200 hours of productive work.
Are there AI coding tools that are safer for SMBs to adopt?
Safety is less about the tool and more about how you configure and govern it. That said, some tools offer features that reduce risk.
Look for tools that offer on-premises or private-cloud deployment, so your code never leaves your infrastructure. Look for tools that allow you to disable data collection or opt out of model training. Look for tools that integrate with your existing security stack (single sign-on, role-based access control, audit logging).
GitHub Copilot for Business, for example, offers enterprise controls, data residency options, and a commitment not to use your code for model training. AWS CodeWhisperer offers similar features. These tools cost more than free or consumer-tier options, but the additional control may be worth it if you handle sensitive data or operate in a regulated industry.
Open-source models that you host yourself (like Code Llama or StarCoder) give you complete control over data flow and usage, but they require more technical expertise to deploy and maintain. For most SMBs, that trade-off is not favorable unless you already have in-house AI expertise.
No tool is risk-free. Even the most secure AI assistant can make mistakes if given vague instructions or excessive permissions. The tool is not the solution. Governance is the solution.
What questions should you ask your IT team about AI coding tools?
If you do not have an internal IT leader, these are the questions to ask your MSP or outsourced CTO.
First, which AI coding tools are our developers currently using? You cannot govern what you do not know about. Get a list.
Second, do we have a written policy that specifies which tools are approved and how they can be used? If the answer is no, creating one is your first priority.
Third, what permissions do these AI tools have? Can they commit code directly to production? Can they access customer data? Can they modify infrastructure configurations?
Fourth, how do we audit AI tool usage? Are there logs? Are they reviewed? How quickly would we detect if a tool caused a problem?
Fifth, what happens if an AI tool makes a mistake? Do we have backups? Can we restore quickly? Do we have a communication plan for clients if the mistake affects them?
Sixth, are we compliant with our contractual and regulatory obligations when we use these tools? If you have client contracts that specify data handling requirements, or if you operate under HIPAA, CMMC, SOC 2, or other frameworks, make sure your AI usage does not create a compliance gap.
These questions do not require deep technical knowledge to ask. They do require a willingness to treat AI as a business decision, not just a developer productivity tool.
When should an SMB prohibit AI coding tools entirely?
Prohibition is rarely the right answer, but there are scenarios where the risk outweighs the benefit.
If your business handles highly regulated data and you cannot configure AI tools to meet compliance requirements, prohibition may be necessary until compliant options become available. For example, if you are a HIPAA-covered entity and the only AI tools your developers want to use send data to third-party servers with no business associate agreement, you have a compliance problem that cannot be solved by policy alone.
If your team lacks the expertise to implement and enforce guardrails, prohibition may be a temporary measure while you build that capability. Allowing unconstrained AI use in a high-risk environment is worse than waiting six months to do it safely.
If your clients have contractual requirements that explicitly prohibit AI-generated code or third-party data processing, you must honor those agreements. Violating a client contract to save developer time is a losing trade.
In most cases, the right path is not prohibition but controlled adoption. Identify low-risk use cases (documentation, test generation, code suggestions that are always reviewed), implement guardrails, train your team, and expand usage as you gain confidence.
What role should your MSP play in governing AI coding tools?
If you work with a managed service provider like TC3, this is a conversation worth having. An MSP can help you inventory the AI tools in use, draft acceptable-use policies, configure technical controls (like branch protection or secret scanning), and audit compliance.
An MSP can also provide a second opinion on vendor claims. When an AI tool vendor promises that your data is secure or that their model never retains your code, your MSP can evaluate those claims, review the vendor’s security documentation, and identify gaps.
For SMBs without a full-time IT leader, the MSP often serves as the de facto CTO. That role includes helping you assess AI adoption security risks, not just responding after something goes wrong.
The key is to engage before you have a problem. Asking your MSP to review your AI governance after an incident is better than nothing, but asking them to help you build governance before adoption is better still.
Frequently Asked Questions
Can AI coding tools delete files without permission?
AI coding tools operate with the same permissions as the user invoking them. If the developer has permission to delete files, the AI can delete files. There is no separate approval gate unless you configure one through your version control or access management system.
Are free AI coding tools safe for business use?
Free AI coding tools may send your code to external servers for processing, retain it for model training, or lack enterprise security controls. For business use, especially with sensitive or proprietary code, paid enterprise versions with data residency guarantees and no-training clauses are safer.
Do I need a policy if only one developer uses AI tools?
Yes. A single developer with an AI tool has the same potential to cause damage as a team of developers. A written policy establishes expectations, creates accountability, and gives you a baseline to audit against. It also protects you if that developer leaves and a replacement starts using different tools.
What should I do if I discover my team is using unapproved AI tools?
Start with a conversation, not discipline. Understand why the team chose that tool, what problem they are solving, and whether an approved alternative exists. Then decide whether to approve the tool with guardrails, switch to a compliant alternative, or prohibit it. Document the decision and update your policy.
How often should we review our AI coding tool policies?
Review policies at least annually, or whenever you adopt a new tool, experience an incident, or face a regulatory audit. AI technology evolves quickly, and a policy that was appropriate six months ago may be outdated today.
Keep reading
Sources
Source: Claude Code Agent Allegedly Deletes 48,000 Files in 103 Seconds