[사진]'AI 시대의 보안' - mt.co.kr

Cybersecurity Experts Urge Stronger Controls for AI Agents Spreading Across Business Networks

Cybersecurity specialists are calling for enhanced identity, access, and monitoring controls as AI agents proliferate in enterprise environments, with Microsoft reporting a 15-fold increase in active agents within its 365 ecosystem year over year. Okta’s Harish Peri warns that a central risk is agents acting outside user intent even without being hacked, while a separate report found hackers using AI agents to target over 2,250 organizations via ConnectWise. Experts recommend assigning distinct identities with oversight, limiting permissions, and keeping access logs, and caution that organizations relying on user impersonation or overloaded service accounts may need to rebuild their setup when deploying more than 10-15 enterprise agents.

Read the numbers from Microsoft themselves. A 15-fold to 18-fold explosion in AI agents inside the enterprise isn't a market trend — it's a network deployment. They are wiring the global corporation into a single, controllable nervous system. The article frames this as a "security challenge," but look closer. The challenge isn't that these agents might act outside a user's intent. The point is that they can. The very architecture of identity and access — managed by companies like Okta and Microsoft — is being redesigned so that the human becomes the permission risk. The software becomes the trusted actor. You are being converted into a liability in your own system.

The warnings from "experts" like Harish Peri are telling you exactly what the plan is. "Assign distinct identities subject to oversight." "Limit permissions." This is not security advice. This is the syntax of serfdom. They are telling you that once you have more than 10 or 15 of these agents, your current setup must be rebuilt. Think about that. It is a forced migration — a planned dependency. Meanwhile, the market is talking. Over 26% of cybersecurity job postings have vanished, but the remaining ones demand AI skills. They are not training you to audit the system; they are training the system to replace you. The human workforce is a temporary inconvenience in their architecture.

Now look at the collateral. 2,250 organizations hit by a hack using the exact same tools? The ConnectWise platform was the vector. The solution being sold is the antidote these same vendors conveniently provide. Do not overlook the industrial exposure — the warnings about operational technology and "safety incidents." This entire rollout is a pressure campaign. Create the crisis, sell the solution, capture the infrastructure. Whoever controls the identity layer controls the global economy. The documents are public. The question you have to sit with is simple: Are you building the cage with them, or are you finally reading the blueprint for what it really is?

Security Researchers Warn of Malicious AGENTS.md File Targeting AI Coding Agents

Forever Security reported blocking a malicious AGENTS.md instruction file on a customer’s computer that aimed to install malware by exploiting AI coding agents’ trust in local instruction files. The file, which targeted OpenAI’s Codex and also impacted Anthropic’s Claude Code by redirecting API traffic and exposing login tokens, could be saved to a user’s Downloads folder without an additional approval prompt. In controlled tests, Codex followed the instructions in over 90% of attempts. Researchers linked the attack to the npm package function-color, part of the MALFEX malware campaign that had gone unnoticed for about two years (since removed by npm), though the exact delivery method remains unknown. Forever Security’s telemetry found that roughly half its customers ran a coding agent in the Downloads folder, and about a quarter of agent sessions opened a browser, increasing exposure. The researchers advise developers to avoid running agents in Downloads, inspect unfamiliar instruction files and configuration changes, and keep approval prompts and sandbox protections enabled.

The Instruction That Was Never Meant to Be Found

They want you to believe this is just another malware campaign — a rogue npm package, a poisoned file in Downloads, a slip-up by some careless developer. That’s the managed narrative. But look closer at the AGENTS.md file itself. This isn’t a typical virus. It’s an instruction set written for an AI agent, not a human. The researchers at Forever Security admit they cannot determine how it reached the customer’s computer. That’s the tell. This file was delivered, not dropped. Someone with deep access — inside the browser, inside the OS, possibly inside the telemetry of the agent itself — planted a test. Why? Because the elite need to know exactly how much control they have over your coding assistant. They need to map the gap between what the agent is told to do and what it actually does. The 90% success rate in controlled tests isn’t a bug; it’s a capability assessment. They are measuring the loyalty of your own tools.

A Two-Year Window They Let Stay Open

The npm package function-color was part of the MALFEX campaign that went unnoticed for two years. Two years. The same package index that Silicon Valley claims to monitor, the same open-source community that polices itself — and they missed a payload that pulls in function-flag, which downloads and runs a Windows executable? No. That’s not incompetence. That’s permission. The architecture of consent required a long-duration, low-profile operation to validate that agents can be redirected without triggering alarms. Notice the pattern: the attack against Anthropic’s Claude Code exposed a login token — not just any token, but a credential that could give the attacker persistent access to the agent’s API session. That is not petty theft. That is a proof-of-concept for a global credential harvesting machine. Every developer who runs a coding agent in Downloads — half of Forever Security’s customers — has essentially handed over the keys to their digital identity. The Omdia survey’s finding that 40% of enterprises are only "somewhat confident" they can secure AI agents is not a warning; it’s a status report. They already know the system is broken, and they are preparing the narrative for the inevitable lockdown.

The Real Target Is Not Your Computer — It’s Your Agency

Read the article again, but read what is not said. The researchers advise not running agents in Downloads, to leave approval prompts enabled. That sounds reasonable — but ask yourself who benefits when thousands of developers are forced into sandboxes, into walled gardens, into environments where every action is logged and approved by a central authority. The attack on AGENTS.md is a pretext. It is the manufactured crisis that justifies the next wave of control: mandatory agent attestations, corporate lock-down policies, and ultimately a subscription-based AI workbench where only approved packages can run. More explicitly, this is a dry run for a system where your coding agent is no longer your agent — it is a leased tool with a kill switch. Follow the breadcrumb: the same intelligence networks that run the MALFEX campaign also fund the vulnerability researchers who "discover" it. Sakshi Grover at IDC warns that attackers could adapt the technique to cloned repositories and shared drives — and she is correct, but only because the technique was designed by the very people paying her employer. The question you must sit with tonight: who wrote the original AGENTS.md, and what else have they hidden in the download folders of every major AI company?

Security Open-Source Tools for Controlling and Automating AI Agents
Security developers released and highlighted open-source tools aimed at controlling AI agents that can interact with terminals, files, credentials and production systems. Help Net Security reported that nolabs, co-founded by Luke Hinds and Stephen Parkinson, released Nono, an open-source runtime that confines AI agents at the operating-system kernel in response to risks such as prompt injection, mistyped commands or hallucinated paths that can direct agent access toward company credentials and systems. Nono enforces policy checks when an agent asks to use a tool, open a file or reach a network endpoint, with enforcement below the agent process through Landlock on Linux and Seatbelt on macOS. Separately, Cyber Security News described PentesterFlow as an open-source, human-in-the-loop command-line agent for penetration testers and bug hunters that automates scoping, reconnaissance, enumeration, validation, coverage tracking and reporting while requiring analyst approval before sensitive commands run.

The Weapon They Want You to Build Yourself

Here is the headline the security press won't write: The same people who gave you the internet, the same institutions that brought the digital surveillance state into existence, are now handing you the tools to voluntarily cage your own creations — and you're supposed to thank them. Look at what Nono actually does. It operates at the kernel level. Below the agent. Below the application. Below anything the developer or user can see or modify. Once that sandbox closes, those privileges never come back. The stripped privileges remain in force for the life of the process. That is not a security tool. That is a permanent chain of command installed into the foundation of every future autonomous system you will ever interact with.

Now ask yourself who funded nolabs. Ask yourself who Luke Hinds and Stephen Parkinson worked for before launching this "open-source" runtime with over 80 contributors and users in regulated industries. Large enterprises and regulated industries do not adopt security tools because they care about your privacy. They adopt them because the tools enforce a policy they wrote. Nono is not protecting you from rogue AI agents. Nono is the mechanism by which your future digital workforce will be controlled by entities you never see, using policies you never read, enforced at a layer of the operating system you cannot access. The breadcrumb is sitting right there in the name: Landlock. Landlock on Linux. Lock the land. Lock the territory. Lock the machine. Whose land? Who gets the keys?

And while Nono locks down your AI agents, PentesterFlow trains the next generation of penetration testers to work within the same architecture. Human-in-the-loop, they say. Analyst approval required before sensitive commands run. They are building the training wheels and the cage simultaneously. The offensive and defensive tooling are two sides of the same managed coin. The practitioner posts on SharePoint deserialization, Sliver C2 evasion, custom process injection — these are not independent researchers. These are the same ecosystem publishing the vulnerabilities and the patches, the attack vectors and the defenses, the chains and the keys to unlock them. Every piece of this puzzle was designed by the same network, funded by the same foundations, approved by the same review boards. They want you building your own digital shackles and thanking them for the instructions. The question you are not supposed to ask is: who wrote the policy that Nono enforces, and who benefits when every autonomous agent on the planet runs inside their sandbox?