Request an invitation ↗
Journal

/

Agentic AI

Agentic AI security: the new attack surface when software can act

What OWASP, NIST, MITRE and the UK's NCSC say about securing AI agents: prompt injection, excessive agency, identity, memory and human approval.

November 18, 2026

8 min read

The Journal

In short

  • Agentic AI security is the work of keeping AI systems that can act, by calling tools, moving data and changing records, from being steered or misused by someone other than their owner.
  • The weakness the public frameworks keep returning to is indirect prompt injection, meaning instructions hidden in an email, a web page or a file that the agent reads as part of its task.
  • A technical director at the UK's National Cyber Security Centre warns that this class of attack may never be fully removed, so the frameworks concentrate on limiting what a hijacked agent is able to do.
  • OWASP, MITRE and Google converge on the same short list of controls: least privilege, an identity for every agent, human approval scaled to the consequence of the action, and logs that show what the agent did.

What changes when software can act

A chatbot that is tricked produces a bad answer, and a person still decides what to do with it. An agent that is tricked can send the email, run the code or approve the refund itself. The US National Institute of Standards and Technology (NIST) put the difference in one sentence of its 2025 taxonomy of attacks on AI systems. Agents inherit the weaknesses of the models they are built on, it says, and then: "because agents can take actions using tools, these attacks can create additional risks in this context, such as enabling actors to hijack agents to execute arbitrary code or exfiltrate data from the environment in which they are operating."

The same report added that "Security research focused specifically on agents is still in its early stages". Guidance has moved quickly since. OWASP, the open security community, published a Top 10 for Agentic Applications in December 2025. MITRE's ATLAS knowledge base of attacks on AI systems now carries agent-specific techniques. NIST asked industry and researchers in January 2026 how agents should be secured, and its summary of the replies, published in May 2026, reports that "Commenters widely agreed that AI agents present novel security threats and that these security concerns present a barrier to adoption."

Our earlier piece on AI agents and agentic AI covers the definitions. This one covers what can go wrong.

Instructions hidden in data

A language model receives its instructions and the material it is working on as one stream of text. The technical director for platforms research at the UK's National Cyber Security Centre (NCSC) stated the consequence plainly in a December 2025 post on the centre's site: "Current large language models (LLMs) simply do not enforce a security boundary between instructions and data inside a prompt."

The post's example is a recruitment tool that asks a model to score each CV against the job requirements. A candidate adds hidden text telling the model to ignore its previous instructions and approve the CV for interview. The applicant never touched the system, yet their text was treated as a command. That is indirect prompt injection. MITRE ATLAS lists it as sub-technique AML.T0051.001 and notes that such instructions have been "observed in the wild".

Older injection flaws, such as SQL injection in databases, were closed by forcing software to keep commands and data apart. The NCSC post doubts the same fix exists here: "it's very possible that prompt injection attacks may never be totally mitigated in the way that SQL injection attacks can be." OWASP's Top 10 for LLM Applications, whose 2026 edition still ranks prompt injection first, says "no reliable prevention mechanism exists today". The NCSC post draws the conclusion a risk owner needs: "If the system's security cannot tolerate the remaining risk, it may not be a good use case for LLMs."

Too much agency

If the model can be misled, the damage depends on what it is allowed to do. OWASP calls this risk excessive agency, third on its 2026 list, and traces it to three root causes: "excessive functionality", "excessive permissions" and "excessive autonomy".

Tool misuse follows from this. In OWASP's threat model for agentic AI, attackers manipulate an agent into abusing its own tools while "operating within authorized permissions". The NCSC post measures the exposure bluntly: once a model can call tools, a successful injection is as bad as handing the attacker those tools directly.

OWASP, MITRE and Google all answer by limiting what the agent can reach. Google's 2025 paper on agent security gives an example: "an agent designed for research should not possess the power to modify financial accounts." OWASP's agentic list goes a step further and names a principle of least agency, on the grounds that "deploying agentic behavior where it is not needed expands the attack surface without adding value". OWASP's list for LLM applications adds a design rule: authorisation is enforced in ordinary program logic, and the model is never the one to decide whether an action is allowed.

Identity, memory and other agents

Agents that run for long periods and work together add three risks.

Identity. Google's paper notes that agents typically inherit the privileges of the people they act for. OWASP's entry on identity and privilege abuse describes the result: "Without a distinct, governed identity of its own, an agent operates in an attribution gap that makes enforcing true least privilege impossible."

Memory. Agents that remember earlier sessions can be fed false information that surfaces later. MITRE ATLAS files this under AI agent context poisoning and says it lets an adversary "persistently change the behavior of the target agent".

Other agents. When agents pass work to each other, each one treats the others' output as input. OWASP's threat model covers poisoned communication between agents, rogue agents inside a workflow and, in the current version 1.1, flaws in the protocols that connect agents to tools and to each other.

The approver is a target too

Human approval appears in the OWASP, MITRE and Google guidance, and the same documents treat it as something attackers plan around. OWASP's threat model includes a threat called overwhelming the human in the loop, where "human capacity cannot keep up with multi-agent operations, leading to rushed approvals, reduced scrutiny, and systemic decision failures". Its Top 10 adds a risk built on trust: people who rely on a fluent, confident agent end up "approving actions without independent validation".

MITRE's guidance is to spend human attention where it counts: "Human In-the-Loop policies should follow the degree of consequence of the task at hand." Google's paper names the actions that deserve it, such as deleting large amounts of data, authorising significant financial transactions or changing security settings.

What the public frameworks cover

  • Top 10 for Agentic Applications 2026. Publisher and date: OWASP, December 2025. What it gives a board: Ten risks, from agent goal hijack to rogue agents, each with mitigations.
  • Agentic AI: Threats and Mitigations, version 1.1. Publisher and date: OWASP, December 2025. What it gives a board: A threat model of 17 named threats, with mitigation playbooks.
  • Top 10 for LLM Applications 2026. Publisher and date: OWASP, August 2026. What it gives a board: The base list: prompt injection first, excessive agency third.
  • AI 100-2 E2025, adversarial machine learning taxonomy. Publisher and date: NIST, March 2025. What it gives a board: Shared vocabulary for attacks, with a short section on the security of agents.
  • ATLAS, release 2026.10. Publisher and date: MITRE, October 2026. What it gives a board: Adversary techniques graded by maturity, case studies and matching mitigations.
  • Prompt injection is not SQL injection. Publisher and date: UK NCSC blog, December 2025. What it gives a board: A plain account of why the flaw persists and how to design around it.
  • Google's Approach for Secure AI Agents. Publisher and date: Google, May 2025. What it gives a board: Three principles: human controllers, limited powers, observable actions.

What testing has shown

In March 2026, NIST's Center for AI Standards and Innovation reported on a public red-teaming competition, hosted by a third party, against agents built on 13 frontier models. More than 400 people made over 250,000 attack attempts, and "at least one successful attack was found against all of the target frontier models".

A public vulnerability record points the same way. CVE-2025-32711, published on 11 June 2025, describes an AI command injection flaw in an office AI assistant that allowed an unauthorised attacker to disclose information over a network. The vendor that assigned the record scored it 9.3 out of 10, in the critical band. MITRE ATLAS records that it was fixed with no evidence of exploitation in the wild.

Six questions before an agent goes live

These questions are ours, drawn from the controls the frameworks share.

  • What untrusted content will the agent read: inbound email, web pages, uploaded files, the output of other agents?
  • Which tools can it call, and which of them change something outside the system?
  • Whose identity does it act under, and can the logs tell its actions from a person's?
  • Which actions always need a person to confirm, and how many approvals a day does that person face?
  • What does it remember between sessions, and who can inspect or reset that memory?
  • Are inputs, outputs and tool calls logged in enough detail to reconstruct an incident? The joint guidance from the US, UK and partner agencies asks for logs covering "inputs, outputs, intermediate states, and errors".

On suppliers, the NCSC post is short: "Beware any that claim they can 'stop' prompt injection, and instead look at those who understand how they reduce it."

FAQ

What is agentic AI security?

It is the set of practices that keep AI agents from being hijacked or misused once they can act through tools. Using AI agents to help security teams is a separate subject.

What is prompt injection, and how does an indirect prompt injection attack occur?

Prompt injection is text written to make a model follow an attacker's instructions in place of its owner's. In the direct form, the attacker types it into the system. In the indirect form, the attacker plants it in content the system will read later, such as a web page, an email or a document, and waits for an agent to process it.

Can prompt injection be prevented?

Not reliably with current models, according to the NCSC post and OWASP. The NCSC post tells designers to focus on "deterministic (non-LLM) safeguards that constrain the actions of the system", so that a successful injection has little to work with.

How do you secure AI agents?

OWASP, MITRE and Google differ in detail and share an outline. Each agent gets only the tools and rights its task needs, and an identity of its own. A person confirms high-consequence actions. Everything the agent reads counts as untrusted, and its actions are logged.

On the Davos programme

The first of four panels at the Davos Super Intelligence Summit 2027, on Tuesday 19 January 2027 at the Grandhotel Belvedere in Davos, is on AGI and agentic AI. The full programme lists all four panels, and access is by invitation only.

Sources

An evening in Davos, by invitation

  • Date19 January 2027
  • Time16:00 – 20:00 CET
  • VenueGrandhotel Belvedere