The Control File
AI Cybersecurity, Explained. Plain-language security for everyone, with practical steps for the organizations that answer for AI at work.
Suppose you set up an AI assistant to do one helpful thing: read your incoming email each morning and give you a summary. To connect it, you click through a permissions screen. You click “Allow” because everyone clicks “Allow.”
The tool you chose can read your mail. It can also send mail, because the company that built the connector bundled both. You never intended the assistant to send anything. But it can. And one morning, an email arrives that knows it can.
This exact situation is the worked example in the OWASP Top 10 for Large Language Model Applications, the most widely used public list of AI application security risks. In OWASP’s version, a malicious email arrives containing hidden instructions. The assistant reads it, follows it, searches the inbox for sensitive information and forwards it to the attacker (OWASP, 2025). Nothing was hacked in the traditional sense. The assistant simply had a key it never needed.
Here is my proposition: with AI agents, the permissions screen you click through is the security system. The damage a fooled agent can do is decided long before the attack, by whoever decided what it was allowed to touch.
The receipt: a name for the gap
OWASP calls this risk excessive agency. It describes the vulnerability that allows an AI system to take damaging actions in response to output that is unexpected, ambiguous or manipulated, and it stresses that the cause does not matter: the trigger could be the model making something up, a poorly performing model, or a hidden instruction planted by an attacker (OWASP, 2025). If the system has the power to do harm, sooner or later something will steer it there.
OWASP breaks excessive agency into three root causes, and they are worth memorizing because they apply to every AI agent you will ever meet:
- Excessive functionality: the agent has tools it does not need, such as the ability to send when it only needs to read.
- Excessive permissions: the tools it has reach further than the job requires, such as access to the whole inbox, or the whole database, rather than one folder or one table.
- Excessive autonomy: it can take high-impact actions without anyone checking first.
Each one maps to a fix. In OWASP’s email example, the attack is stopped by giving the assistant a read-only tool, connecting it with a read-only permission, and requiring the user to review and send any drafted email themselves (OWASP, 2025). Any one of those three would have helped. All three together leave the hidden instruction with nothing to work with.
Key idea: excessive agency
Excessive agency is when an AI system has more tools, more access or more freedom to act than its job requires. It matters because AI agents can be confused or manipulated, and the damage a confused agent can do is set by what it is allowed to touch. Shrinking an agent’s tools, access and freedom to act is the most reliable way to protect the people whose data and money it handles.
Who benefits from the extra keys, and who pays
Bundled permissions are convenient for the people building and selling these tools: one connection, one screen, fewer support tickets. The bill arrives elsewhere. It lands on the person whose inbox is searched, the customer whose data is forwarded, and the small business that has to explain it.
Why the oldest rule in security struggles here
Least privilege, giving any user or system only the access it needs for its task, is one of the oldest ideas in security. Agents strain it in two ways.
First, an agent’s job is often open-ended. The US National Institute of Standards and Technology (NIST) put this plainly in a February 2026 draft concept paper on agent identity and authorization. Among the questions it asked industry: how do we establish “least privilege” for an agent, “especially when its required actions might not be fully predictable when deployed?” (Booth et al., 2026). It also asked how to handle agents acting “on behalf of” a person, and how to tie an agent’s identity to a human’s so that approvals and logs point to someone accountable. That a standards body is still asking these questions tells you the answers are not settled.
Second, agents act under a trusted identity, so their misuse can look legitimate. Joint guidance published in May 2026 by six cyber security agencies from Australia, the United States, Canada, New Zealand and the United Kingdom describes a procurement agent with more access than it needs. An attacker who gets into its workflow inherits those privileges and can change contracts and approve payments. Because everything runs under the agent’s trusted identity, the audit logs “appear legitimate and delay detection” (ASD’s ACSC et al., 2026, p. 8). Security specialists call this a confused deputy: a helper with real authority, tricked into using it for someone else.
The best argument against me
The strongest objection is practical. Agents are useful precisely because they can act. Strip away their tools and you are back to a chatbot that writes drafts you have to send yourself. And as NIST’s own question admits, you cannot always predict in advance what an agent will need.
That is true, and it is why the answer is not “no access.” It is access matched to the job, with a human yes on the actions that can’t be undone.
Three questions that shrink the keyring
You do not need to solve agent identity to make progress. For any AI agent, at home or at work, ask three questions that mirror OWASP’s three root causes.
1. What does it need to do? List the actions, then remove every tool that is not on the list. OWASP specifically warns against open-ended tools, such as the ability to run any command, when a narrow tool would do (OWASP, 2025).
2. What does it need to reach, and for how long? Read-only where possible, one folder rather than the whole drive, and access that expires. The UK’s National Cyber Security Centre puts it simply: “give agents only the minimum access they need, for the shortest time required” (NCSC, 2026).
3. Which actions need a human yes? Sending, paying, deleting and sharing outside the organization are the usual candidates. And enforce that rule outside the AI: OWASP’s advice is to “implement authorization in downstream systems rather than relying on an LLM” to decide what is allowed (OWASP, 2025). The bank should refuse an unapproved payment even if the agent asks nicely.
Finally, record what each agent did and under whose authority. OWASP notes that logging and rate limits do not prevent excessive agency, but they limit the damage and shorten the time before someone notices (OWASP, 2025).
The verdict
On the guidance we reviewed, from OWASP, NIST and six national cyber agencies, the most reliable protection against a manipulated agent is not a smarter model. It is a smaller keyring. An agent that can only read cannot forward your inbox, however cleverly it is tricked.
What we still don’t know
The scenarios in this post come from security guidance, not from incident reports; they show how harm can happen, not how often it has. We found no reliable public figure for how many deployed AI agents are over-privileged. NIST’s paper is a draft that asks questions rather than setting requirements, and its comment period closed in April 2026; we have not seen a final project description. And least privilege is not a cure: an agent with minimal access can still make mistakes within that access. It simply keeps those mistakes smaller.
Which leaves the question nobody has answered yet: when an agent acts “on behalf of” you, and something goes wrong, whose name ends up in the log?
Questions you can ask
- Before you connect a personal AI assistant: What exactly does this permission screen let it do, and can I give it less?
- If a company’s agent acts on your account: What can it do without a person checking, and how do I undo what it did?
- If you deploy agents: For each agent, can we show its tools, its permissions, and the actions that need human approval, as they are today?
- If you approve AI projects: Who reviews an agent’s access when its tools change, and who is accountable for what it does?
Go further. Operational AI Readiness is FairByDesign’s free field manual for AI that is already live. Chapter 7, Security Controls for LLM Apps and Agent Workflows, turns least privilege into concrete tool and permission boundaries, and Chapter 9 covers approval gates for high-impact actions. Domain 7 of the free readiness assessment checks the same boundaries. Read the book free.
Help keep this free and independent. We explain AI security in plain language, without vendor sponsors and without a product to sell you, and we keep the field manual free for anyone who needs it. That only works because readers fund it. If this helped, please give once ($5, $10, $25, $50 or any amount), give monthly ($10 or $25), or become a patron: fairbydesign.org/patrons-2. Support never buys editorial control.
References
Australian Signals Directorate’s Australian Cyber Security Centre, Cybersecurity and Infrastructure Security Agency, National Security Agency, Canadian Centre for Cyber Security, New Zealand National Cyber Security Centre, & United Kingdom National Cyber Security Centre. (2026, May 1). Careful adoption of agentic AI services [Joint guidance]. https://www.ncsc.govt.nz/assets/guidance/Documents/Careful-adoption-of-agentic-AI-services_FINAL.pdf
Bakos, V. (2026). Operational AI readiness: A practitioner guide to governing AI that is already live. FairByDesign. https://fairbydesign.org/operational-ai-readiness/
Booth, H., Fisher, B., Galluzzo, R., & Roberts, J. (2026, February). Accelerating the adoption of software and artificial intelligence agent identity and authorization [Draft concept paper]. National Cybersecurity Center of Excellence, National Institute of Standards and Technology. https://www.nccoe.nist.gov/sites/default/files/2026-02/accelerating-the-adoption-of-software-and-ai-agent-identity-and-authorization-concept-paper.pdf
National Cyber Security Centre. (2026, May 15). Thinking carefully before adopting agentic AI [Blog post]. https://www.ncsc.gov.uk/blogs/thinking-carefully-before-adopting-agentic-ai
OWASP GenAI Security Project. (2025). LLM06:2025 Excessive agency. OWASP Top 10 for LLM Applications. https://genai.owasp.org/llmrisk/llm062025-excessive-agency/




Leave a Reply
You must be logged in to post a comment.