An AI agent reads content whose origin it does not control, chooses the next action itself and executes it with real permissions in the information system. AI agent security is shifted as a result: the examination is less about what the agent is supposed to do than about what it could do if the model made a mistake or were manipulated.
It therefore consists of limiting what the agent can reach, whatever the model decides: permissions reduced to the task, a closed list of tools, a justified reading scope, content read treated as untrusted, a trace of every call and a tested shutdown. On each of these points, a chief information security officer (CISO) can ask a question before authorising go-live. This grid is a proposed method, based on public sources consulted on 12 November 2026: the OWASP Top 10 for LLM and generative AI applications (“2025 Top 10 Risk & Mitigations for LLMs and Gen AI Apps”, genai.owasp.org) and the Model Context Protocol documentation (modelcontextprotocol.io).
What could an attacker do with the agent’s permissions?
If the model obeys the attacker, the instruction ceases to be a protection, and the permissions of the account used limit the damage.
OWASP names this risk “Excessive Agency” (LLM06:2025) and gives it three origins: excessive functionality, permissions or autonomy. One of its examples describes an extension designed to read data, which connects to the database with an identity that is also authorised to update, insert and delete.
The CISO asks for the record of actual permissions, application by application, and reads it as a worst-case scenario. They look in it for an account specific to the agent, time-limited tokens held by the code outside the model’s context and, when the agent acts for a person, permissions that do not exceed that person’s own. The same entry cites the opposite case: a privileged account that opens the files of all users. This is least privilege, granted session by session, as presented in our entry on zero trust.
Is each tool opened to the agent necessary for its task?
The CISO asks for the list of tools as deployed, with, for each one, what it reads, what it writes and what cannot be undone.
OWASP gives examples of a list that overflows. A third-party connector, chosen to read documents, can also modify and delete them. A tool tried out during development, then dropped, remains accessible to the agent. A tool that is open by nature, such as executing a command or fetching an arbitrary web address, is replaced by a tool limited to one operation.
The control is placed in the application being called. OWASP speaks of complete mediation: authorisation is checked by the downstream system, instead of being left to the model’s judgement.
A third-party tool server is a software dependency, whose provenance and updates follow the controls in our article on application security. For a local server connected via the MCP protocol, the security documentation requires a client application that offers one-click installation to display the exact command and obtain explicit consent before executing it. The design of tools and permissions is described on our AI agents page.
What does the agent read, and how can what it has read come back out?
What the agent reads enters its context and may end up in what it produces next: an answer, the parameters of a tool call, a note kept in memory. OWASP deals with disclosure through outputs under the heading “Sensitive Information Disclosure” (LLM02:2025).
The CISO asks for a record of the sources (document repositories, messaging systems, databases, web pages, memory kept from one session to the next), indicating, for each one, whether the agent reads with its own permissions or with those of the person being served. A shared agent that reads with a broad account can show someone what they would not have been entitled to open.
They also ask for the record of outputs: sending a message, calling an external service, writing to a shared space, a link inserted in a displayed answer. The Model Context Protocol specification (“tools” page, read on 12 November 2026) recommends that client applications show the user the parameters of a call before sending it to the server, to avoid data exfiltration. A model operated by a third party counts among these outputs: what enters the context is transmitted to that third party, under the terms of its contract.
What becomes of AI agent security when content it reads gives orders?
OWASP describes prompt injection under number LLM01:2025 and distinguishes an indirect form, which occurs when the model accepts input from external sources, such as websites or files. For an agent, an email, an attachment or the result of a tool are such sources.
The same entry sets a limit: given the way models work, it is unclear whether there are fool-proof methods of prevention. The CISO therefore reasons from the assumption that the injection succeeds, and looks for what then stops the consequence.
Permissions, the list of tools and the record of outputs answer this in part. OWASP adds human approval for high-risk actions, segregation of external content and regular attack testing, in which the model is treated as an untrusted user. The file handed to the CISO contains injection scenarios played against the agent as it will be deployed, with the action attempted and the barrier that stopped it. When several agents cooperate, the path of booby-trapped content depends on the mode of agent orchestration.
Will it be possible to reconstruct what the agent did, call by call?
After an incident, the investigation needs, for each execution, what triggered it, the content read, each tool call with its parameters and its result, and the approvals requested. The agent’s own account makes its actions recognisable in the applications’ logs.
Traces are written out of the agent’s reach, so it can neither modify nor erase them. They feed detection: a blocked out-of-scope action or an unusual volume of calls triggers an alert. Their retention period is set with the data protection officer, because personal data may appear in them.
OWASP’s LLM06:2025 entry lists logging and rate limiting of calls among the measures that limit damage without preventing the risk.
How do you shut the agent down, and what remains to be undone?
Shutdown is a technical procedure: disable the agent’s account, revoke its tokens, cut off access to its tools. The file states who can trigger it and what happens to tasks in progress. The designation of supervisors falls under the authorisation given by the committee, the subject of our article on enterprise use cases for AI agents.
Once the agent has been shut down, the traces give the list of its recent actions, and reversible writes are undone one by one. Actions that cannot be undone (a payment, a message sent to a third party, a deletion) justify prior approval limited to those actions.
If the agent served as an entry point for an intrusion, business continuity and recovery fall under the cyber resilience set-up, which must know about the agent and its accounts. The CISO’s approval and reservations can be added to the register of systems kept by AI governance.
A shutdown test run before go-live, timed from the decision to the last tool call, gives the CISO the final piece of their examination.