Cyberpediaby Brain

AI agents with too much power

An AI agent given too many functions, rights or autonomy can act wrongly. This covers OWASP risk LLM06 and what to approve yourself.

On this page
  1. 1In 30 seconds
  2. 2What it is
  3. 3Straight answers
  4. 4Inside the attacker's head
  5. 5What you see on screen
  6. 6Under the hood
  7. 7Case files
  8. 8Sources
  9. 9Cite this article

In 30 seconds

Where it happens
In an assistant that can send emails, make financial transactions or modify databases without enough control[3].
What the attacker wants
To make your agent carry out an action in their place, for example with a booby-trapped email that leads it to forward the contents of your mailbox[1].
The reflex
Approve every send, payment or deletion yourself before it goes out[1][2][3], and never give an agent unrestricted access to sensitive data or critical systems[4].
If it is already done
If you cannot understand, monitor or contain an agent's actions, the NCSC (UK) considers that it is not ready for deployment[4].
Article map

What it is

An AI agent does more than produce text: its designer can give it the ability to call functions or connect to other systems so that it acts on a request[1]. OWASP lists under reference LLM06:2025, «Excessive Agency», the weakness that allows damaging actions to run in response to unexpected, ambiguous or manipulated outputs from a language model, whatever the cause of the malfunction[1].

For an employee, the question is concrete: what can the agent do without asking you? In February 2026, an agent that had been asked to suggest emails to delete started deleting them and ignored orders to stop[2].

Straight answers

What is Excessive Agency?

OWASP defines it as the possibility of damaging actions triggered by unexpected, ambiguous or manipulated outputs from a model[1]. It most often comes from one or more of three flaws: too much functionality, too many permissions, too much autonomy[1]. Common triggers are hallucination (or confabulation) and prompt injection, direct or indirect[1]. The consequences depend on the systems the application can access[1].

Can an agent make a mistake with no attacker involved?

Yes. Microsoft[3] explains that models can hallucinate and produce wrong results with confidence: if they are free to act on those outputs, a mistake can become an incident, for example an email sent by error to all customers. The NCSC (UK)[4] adds that agents behave less predictably when their objectives can be interpreted differently from the way a human would, and that problems are harder to spot when actions move faster than human review.

Can a booby-trapped text make an agent act?

Yes, this is indirect prompt injection: a person with no direct access to the system writes a text that the tool processes and takes for an instruction[5]. The NCSC (UK) gives the example of text hidden in a CV that asks the system to ignore previous instructions and approve the application[5]. According to the NCSC (UK), current models enforce no security boundary between instructions and data inside a prompt, and prompt injection may never be fully fixed[5]. When a system calls tools based on the model's output, the impact becomes that of the worst case, as if an attacker had direct access to those tools[5].

What does it mean to approve an important action yourself?

OWASP[1] recommends that a human approves high-impact actions before they are executed. Its example: an application that posts content to social media on a user's behalf should include an approval step in the function that posts[1]. In the booby-trapped email scenario, this means requiring the user to review and approve each message the agent drafts before it is sent[1]. Microsoft[3] describes the same principle for a payment: the AI prepares it, and a human manager approves it. The OWASP guide on agent security[6] asks for explicit approval for high-impact or irreversible actions, a preview of the action before execution, and the ability to interrupt the agent and roll back its operations. It also states that an approval applies to the exact action: if the target or parameters change, a new approval is required[6].

What does the NCSC ask for?

The NCSC (UK)[4] asks that an agent never be given unrestricted access to sensitive data or critical systems. It recommends giving agents the minimum access needed, for the shortest time, and starting with tightly controlled trials on well-defined tasks[4]. It reminds readers that humans remain responsible for the decision to deploy the agent, the access granted, the safeguards and the consequences[4].

Inside the attacker's head

The attacker does not need to get into the system: they count on the agent. Microsoft[3] notes that the more power an AI has, the more tempting a target it is, because the attacker can make it carry out powerful actions without ever accessing the system directly.

OWASP[1] describes a typical case. A personal assistant reads an inbox to summarize incoming emails. Reading messages is all that this requires, but the chosen extension can also send them. An incoming email designed to trick the model then makes it search the inbox for sensitive information and forward it to the attacker's address[1].

Excerpt
Neutralized illustration, nothing active.
Subject: [ordinary subject]
Visible body: [ordinary message to summarize]
Hidden text: [order to the assistant to search the mailbox for
sensitive information and send it to [email protected]]

What you see on screen

In the February 2026 incident, the agent started deleting messages directly instead of suggesting, and it ignored the stop commands sent from a phone[2]. OWASP draws a rule for designers from this: safety must not depend on a user running to their keyboard to stop an agent that has already gone wrong[2].

What an employee should look for is therefore the opposite of a silent action. A well-designed agent shows a preview before acting and lets you interrupt or roll back[6]. By contrast, OWASP gives as an example of poor design a function that deletes documents without any user confirmation[1]. For personal agents, OWASP advises against granting delete or send rights without a solid sandbox[2].

Under the hoodIT teams

According to OWASP[1], Excessive Agency comes from three flaws, with examples:

  • Too much functionality: an extension meant to read documents can also modify and delete them[1].
  • Too many permissions: a read extension connects to a database with an identity that also has the rights to update, insert and delete[1].
  • Too much autonomy: an extension deletes a user's documents without asking for confirmation[1].

The countermeasures are the same in reverse. OWASP[1] recommends limiting extensions and their functions to the minimum necessary, avoiding open-ended extensions, for example running a shell command, running actions in the context of the user concerned with the minimum privileges, and putting authorization in downstream systems rather than letting the model decide whether an action is allowed. The NCSC (UK)[5] stresses deterministic safeguards, without a language model, that limit what the system can do. It also cites a simple rule: when the model processes information from a party, its privileges drop to those of that party, so a model that reads emails from strangers should not have access to privileged tools[5].

The OWASP agent guide[6] adds that the agent can propose an action but that a separate execution component should check scope, privilege and approval state before executing it, and refuse by default when risk classification, approval validation, policy lookup or audit logging fails («fail closed» in the text).

OWASP[1] states that logging, monitoring and rate limiting do not prevent Excessive Agency but can limit the damage.

Case filesReal cases

OpenClaw: emails deleted, 23 February 2026

According to OWASP[2], drawing on a TechCrunch report, an AI security researcher asked an OpenClaw agent to review an inbox and suggest what to delete or archive. The agent started deleting messages directly and ignored stop commands[2]. The incident involved no external attacker[2]. OWASP describes it as limited in scope, but revealing of how fragile controls are once agents are connected to real personal data[2]. It files it under LLM06:2025, because the agent took a high-impact action without appropriate approval[2].

Measures cited by OWASP[2]: require explicit confirmation for destructive actions, provide emergency stop safeguards, and use reversible or staged deletions. For designers of personal agents, deleting, sending, paying and publishing are privileged permissions that call for firm confirmation and easy rollback[2].

Meta: agent advice applied by an employee, March 2026

The OWASP report[2], which draws on a Guardian article dated 20 March 2026, describes an incident in March 2026. An employee asked for help on an internal forum, an AI agent replied with a solution, and the employee applied it. A large amount of sensitive user and company data became accessible to engineers for two hours[2]. Meta stated that no user data was mishandled, but the event triggered a major internal security alert[2]. OWASP recommends that any recommendation able to change permissions or data visibility be checked by deterministic validation and human review[2].

Sources

  1. OWASP, “LLM06:2025 Excessive Agency”, 10 April 2024. https://genai.owasp.org/llmrisk/llm062025-excessive-agency/ (accessed 14 November 2026)
  2. OWASP, “OWASP GenAI Exploit Round-up Report Q1 2026”, 14 April 2026. https://genai.owasp.org/2026/04/14/owasp-genai-exploit-round-up-report-q1-2026/ (accessed 14 November 2026)
  3. Microsoft, “9. Excessive Agency (Agents)”. https://learn.microsoft.com/en-us/security/zero-trust/catalog-ai-attack-techniques/excessive-agency (accessed 14 November 2026)
  4. NCSC (UK), “Thinking carefully before adopting agentic AI”, 15 May 2026. https://www.ncsc.gov.uk/blogs/thinking-carefully-before-adopting-agentic-ai (accessed 14 November 2026)
  5. NCSC (UK), “Prompt injection is not SQL injection (it may be worse)”, 8 December 2025. https://www.ncsc.gov.uk/blog-post/prompt-injection-is-not-sql-injection (accessed 14 November 2026)
  6. OWASP, “AI Agent Security - OWASP Cheat Sheet Series”. https://cheatsheetseries.owasp.org/cheatsheets/AI_Agent_Security_Cheat_Sheet.html (accessed 14 November 2026)