12 Hours, Unnoticed

The official VS Code extension for Anthropic’s Claude was replaced with a fake version containing a backdoor. For about 12 hours, up to 5,000 developers installed it without realizing.

This occurred through official distribution channels. In other words, the assumption that “official means safe” was shattered.

Do you think this is just an issue for engineers at large companies? No. For small and medium-sized enterprises (SMEs) that are starting to experiment with AI agents, this serves as a structural warning.

What Happened

In 2024, version 2.1.313 of the VS Code extension corresponding to Anthropic’s Claude model was released with malicious code embedded by attackers. The method used was a so-called “supply chain attack.” By simply updating as usual, a backdoor entered the user’s machine.

There are three key points:

• It was through official channels (users had almost no means to distinguish it).
• It was left unattended for 12 hours (it took half a day to detect and respond).
• Up to 5,000 development machines were affected (each machine may contain source code, API keys, and customer data).

If 5,000 development machines were compromised, referencing IBM’s “Cost of a Data Breach Report 2024,” which indicates the average cost of a data breach is approximately $4.88 million (about 730 million yen) per incident, the scale of the damage is beyond imagination. Of course, not all machines will incur actual damage, but considering the possibility of even one API key being extracted, the costs for incident response alone could range from several million to tens of millions of yen.

The Paradox of “Increased Security Costs with AI Usage”

Now, to the main point.

The introduction of AI agents offers clear benefits such as reduced labor costs and automated operations. If an AI tool costing 10,000 yen per month can replace work that would otherwise cost 300,000 yen per month, there is no reason not to implement it. In fact, such cases are increasing.

However, there are often overlooked costs: the operational costs of security.

Incorporating AI agents into operations means increasing the “attack surface” as follows:

Reports indicate that at least nine AI development platforms were compromised simultaneously in October 2024. This is not a coincidence; it reflects a structural change where attackers are beginning to target the “supply chain of AI tools.”

Real Cost Estimates for SMEs—How Much More Will It Increase Monthly?

Enough with the abstract discussion. Let’s concretely estimate the costs for an SME with 20 employees that introduces three AI agents.

[Before AI Agent Implementation]

• Security measures cost: Approximately 20,000 yen per month (antivirus software + firewall)

[Additional Costs Required After AI Agent Implementation]

| Total Additional Costs | Approximately 90,000 yen/month |

The AI tools themselves cost 30,000 yen per month. However, when including security, the total comes to 90,000 yen per month. This means that the security costs are three times the tool costs.

How one perceives this—whether as “expensive” or “cheap”—is straightforward: compare it to the potential damage from a single incident. If an API key is leaked and cloud usage is fraudulently incurred, there are real cases where bills of 500,000 to 1,000,000 yen arrive overnight. If customer data is leaked, the damage to reputation cannot even be quantified in monetary terms. Viewed as “insurance,” 90,000 yen per month is, in fact, the minimum baseline.

So, What Should SMEs Do?

Saying “don’t use AI because there are risks” is not a solution. Not using it will lead to a loss of competitiveness. The issue is “how to use it.”

Here are three immediate actions SMEs should take:

1. Discard the Notion That “Official Means Safe”

What this incident proved is that even official channels cannot be trusted. Establish a rule to wait at least 24 hours before applying updates for extensions and plugins. This alone could have avoided the 12-hour backdoor incident. The cost is zero.

2. Minimize the Permissions Granted to AI Agents

Surprisingly, many companies grant “administrator privileges” to AI agents. The principle should be Least Privilege. If read-only access suffices, do not grant write permissions. Separate API keys by purpose and set usage limits. Simply setting a monthly usage limit alert for cloud expenses can significantly change the damage in case of runaway operations.

3. Decide Who to Contact in Case of an Incident

It is not practical for SMEs to have a full-time security officer. However, whether you have a contact that you can consult within an hour when an incident occurs can change the damage amount tenfold. Ensure you have at least a local IT support organization or a monthly consultation plan with a security vendor (available from around 10,000 yen per month).

The True Cost of “Automation” Is the Hidden Costs of Being Unmanned

The greatest value of AI agents is that they can operate without human intervention. However, because there are no humans, abnormalities can go unnoticed. The 12-hour backdoor was left unattended precisely because of this structure.

Redirect part of the costs saved from automation to monitoring and insurance. This is not merely “defense”; it is an “investment” to use AI confidently for offensive purposes.

SMEs certainly have the potential to compete with large corporations using AI. Their smaller size allows for quicker decision-making and implementation. However, the faster they implement, the quicker they also take on risks. If those risks can be managed for a few tens of thousands of yen per month, there is no reason not to do it.

Judging AI agent usage fees as “cheap” by only looking at the tool costs is akin to deciding to move based solely on rent. Utility costs, insurance, repair costs—only when all hidden costs are considered can one truly assess whether “this investment is worthwhile.”

Start by taking stock of the permissions granted to your company’s AI tools. This will be the first step in determining whether the 90,000 yen insurance is necessary.