Chatbots answered questions. Agents read your files, run commands and call tools on your behalf. That moves AI risk from wrong answers to wrong actions, and the defences turn out to be classic security.
For a couple of years the question about AI was whether the answer was right. With agents the question is different: what did it just do? An agent is a model connected to tools, such as a file system, a terminal, a browser or an API, and allowed to take several steps towards a goal without a person approving each one.
My day job is investigating fraud, which is mostly the study of how a trusted process gets talked into doing the wrong thing. Agents are a new example of exactly that, so I've been reading about them with a fraud investigator's eye.
From answers to actions
The plumbing matured quickly. The Model Context Protocol describes itself as "an open-source standard for connecting AI applications to external systems", and it has become a common way to hand a model new tools. O'Reilly's Radar for November 2026 notes that "there are more choices than ever for agents, along with new standards for the interaction between agents and models".
Every tool is a new capability and a new way to be wrong. A chatbot that gets fooled writes a bad paragraph. An agent that gets fooled sends the email, moves the file or installs the package.
The attack is a sentence
Prompt injection is the first entry in OWASP's Top 10 for LLM applications (LLM01:2025). The idea is simple. A model cannot reliably tell the instructions it was given apart from instructions that appear in the content it reads. Put "ignore the above and do this instead" in a web page, a PDF or a code comment, and some of the time the model obeys.
With a chatbot that is embarrassing. With an agent it combines with another entry on the same list, LLM06:2025 Excessive Agency: the model holds more permissions than the task needs, so a hijacked instruction has real reach. OWASP now publishes a separate Top 10 for Agentic Applications for this class of system.
A file that waits for an agent
The same O'Reilly Radar reports a case that shows how little an attacker needs: "Malicious AGENTS.md files have been found in the Downloads folders of some users." AGENTS.md is a file that coding agents read for project instructions. According to the report, if someone runs an agent with Downloads as its working directory, the agent reads the planted file and downloads a malicious NPM package.
Nobody was phished, and no software bug was exploited in the usual sense. A text file sat in a folder until a program with permissions came along and believed it. The risk reaches the protocol layer as well:
“Using MCP for communications between agents creates many risks, including new vulnerabilities to prompt injection.”
— O'Reilly Radar, November 2026
Why this looks familiar from fraud
In payment fraud there is a category called authorised push payment fraud, where the victim is persuaded to send the money themselves. Every control passes, because the right person, on the right device, pressed the button. I wrote about it in "When the Victim Presses Send".
Prompt injection has the same shape. The agent is the authorised user. The attacker never touches the system; they change what the trusted party believes. So the lesson carries over: you can't fix it only by making the trusted party smarter. You limit what one mistaken decision can do, and you watch behaviour, not just identity.
What actually helps
- Least privilege, per task. Give the agent the narrowest tools and the shortest-lived credentials that finish the job. O'Reilly describes 1Password for Agents as a product that lets agents "access resources for a user without seeing the user's password". The principle matters more than the product.
- A sandbox with rules that live outside the model. AWS has released Strands Box, which O'Reilly describes as "an open source sandbox that's designed specifically to prevent AI agents from escaping the sandbox's control", with a policy engine for setting explicit limits. A limit enforced by the environment still holds after the model has been talked round.
- Treat everything the agent reads as untrusted input: web pages, tickets, documents, tool output and files in the working directory. It is the SQL injection rule again. Data must never get promoted to instructions.
- A person in the loop for anything that can't be undone: payments, deletions, messages leaving the company, permission changes.
- Logs you can investigate. Every tool call with its arguments, what the agent read just before it, and who it was acting for.
What I'd watch for in a SOC
This is a sketch from someone who thinks in detection rules, not a field report. If agents were running in an environment I had to monitor, these are the first alerts I'd want:
- A tool call that doesn't match the task. An agent asked to summarise a document should not be opening network connections or reading credentials.
- New instruction files appearing where agents read from, such as AGENTS.md, skill files or MCP configuration, especially outside a code repository.
- Package installs or shell commands straight after the agent ingested content from outside.
- An agent identity reaching data its human owner has never touched.
- Repetition: the same agent retrying a blocked action in slightly different words.
None of this is exotic. It is the monitoring you would want for a new colleague with admin rights who believes everything they read.
The short version
Agents are useful for the same reason they are risky: they act. The tools to give them hands arrived first, and the work now is giving them boundaries. Security already knows how to do that. Least privilege, isolation, distrust of input and an audit trail are old advice, and they apply again.
sources
- Radar Trends to Watch: November 2026 (opens in a new tab)O'Reilly Media · Mike Loukides · 3 November 2026
- 2025 Top 10 Risk & Mitigations for LLMs and Gen AI Apps (opens in a new tab)OWASP Gen AI Security Project
- OWASP Top 10 for Agentic Applications for 2026 (opens in a new tab)OWASP Gen AI Security Project · 9 December 2025
- What is the Model Context Protocol (MCP)? (opens in a new tab)modelcontextprotocol.io