Prompt injection and Markdown: when an AI agent trusts the wrong text
In brief
A Markdown file does not execute code by itself. It can, however, contain hostile instructions that an AI agent treats as trusted. Risk depends on which content is authorised to guide it and which data and tools it can access.
What to do now
- Define who approves the instructions agents use.
- Have access, available data and authorised operations reviewed.
- Require review before high-risk actions.
Imagine an employee asking an AI assistant to fix a problem in a website’s code. During the task, the AI assistant reads a document suggesting an additional tool should be installed. The suggestion sounds useful, but does the document come from a trusted source? Risk arises if the AI assistant mistakes that guidance for authorised instructions and proceeds with the installation.
That is the central concern behind indirect prompt injection: instructions in external content try to divert an agent from its authorised task. This is not just about the quality of an answer. When AI has operational tools, a wrong decision can affect business systems.
The case reported in October 2026
On 22 October, Forever Security reported finding a malicious AGENTS.md in a customer’s Downloads folder. According to the researchers, it prompted an agent to attempt a dangerous installation, blocked before the malware ran. How the document arrived there was not established. Later laboratory demonstrations should not be treated as additional confirmed incidents. Original research.
The case described comes from Forever Security’s research, not from our own testing. It does not mean that all AI agents are equally vulnerable: risk also depends on their configuration and available permissions.
The same care applies to agent instruction files and skills, packages of instructions and sometimes additional tools downloaded from the web. Whether they come from a repository, a catalogue or another website, their origin, contents and any included scripts need checking before use. The name of the hosting platform alone does not guarantee that they are trustworthy.
A Markdown file is not a virus by definition
Markdown is a text format used for documentation and project guidance. The file does not execute commands by itself. Risk arises when a system interprets its contents as trusted instructions and has permission to act.
A document can be useful to read without being authorised to change the rules of the job. Its filename, professional tone and presence in a project folder do not, by themselves, establish a trusted origin.
Why this matters beyond software development
Imagine an assistant asked to organise sales documents. A note received from outside the company could suggest an action unrelated to that request. An assistant that can only propose a summary presents a different risk from one that can edit files, connect to services or install tools.
A business therefore needs to assess what the agent can access, not just which model it uses. Code, customer documents and service credentials should not become available merely for convenience.
Who decides which files are trusted?
That decision belongs to the organisation, not to the document the agent is reading. OWASP recommends separating external content from authorised instructions, limiting privileges and requiring human approval for high-risk operations. OWASP guidance.
Before introducing an agent into everyday work, clarify:
- Who approves documents that may guide its behaviour?
- Which data and tools does its task actually require?
- Who authorises configuration changes, installations and information transfers?
- How can its actions be reconstructed when something looks wrong?
This is not a call to block every Markdown file. It is a call to distinguish content to consult from authority to obey.
Security should accompany automation
This differs from bot and scraper traffic on a website: the potential weak point here is an assistant authorised to work for the business, influenced by untrusted material.
At BellosatoTech, we assess workflows, access and separation before proposing changes. To understand whether AI tools and digital processes expose important data or operations, start with a technical audit. The goal is not to abandon AI, but to retain its usefulness without handing control to an unknown document.
