# Agentic AI Security: Injection Changes What an Agent Believes. It Should Not Change What It Can Do.

> What agentic AI security covers, why prompt injection is architectural rather than a bug, the lethal trifecta, and why bounding what an agent can do matters more than filtering what it reads.

Source: https://www.kognitos.com/blog/agentic-ai-security/

Published November 19, 2026 by Kognitos.

## TL;DR

Agentic AI security covers protecting systems where an AI agent takes actions rather than producing text. Almost all of the published guidance addresses inputs: filtering prompts, detecting injection, isolating untrusted content. That work is necessary and it is not sufficient, because prompt injection is described by Meta as a fundamental unsolved weakness in all LLMs. The durable control is bounding what a compromised agent is able to do.

**Key Takeaways:** OWASP maps prompt injection to six of the ten categories in its Top 10 for Agentic Applications. The lethal trifecta is private data access, untrusted content exposure, and external communication; any agent with all three can be turned into an exfiltration tool. Blast radius control is the consensus primary defense. Permissions bound where an agent can act; rules bound what it can do there.

## What is agentic AI security?

Agentic AI security is the practice of protecting systems in which an AI agent takes actions across tools and systems, rather than producing text for a person to read.

The distinction matters because the threat model changes. A generative model that is manipulated produces a wrong answer. An agent that is manipulated takes a wrong action, in a real system, with real credentials, and frequently without a human in the path.

OWASP’s GenAI Security Project reflects this shift directly. Its 2025 report on agentic security cataloged plausible threats. The 2026 edition catalogs CVEs, vendor advisories and breach reports across nearly every category of agentic risk. The threats stopped being theoretical inside a year.

## Why prompt injection is architectural, not a bug

Prompt injection is the technique that ties most agentic incidents together, and OWASP maps it to six of the ten categories in its Top 10 for Agentic Applications.

The mechanism is simple and difficult to fix. A language model reads its instructions and its data through the same channel. There is no equivalent to a parameterized query, no syntactic boundary separating “this is a command” from “this is content to process.” When an agent reads a document, a log file, a web page or an email, any instructions embedded in that content arrive in the same stream as the ones you wrote.

Meta’s characterization, quoted widely in the OWASP reporting, is that prompt injection is a fundamental, unsolved weakness in all LLMs.

The practical consequence for a security posture is uncomfortable but clarifying. You should assume a successful injection and design for what happens next.

## The lethal trifecta

The clearest formulation of the risk, from researcher Simon Willison and now cited throughout the OWASP material, is that three properties together create structural danger:

- **Access to private data.** The agent can read things worth stealing.
- **Exposure to untrusted content.** The agent processes input an attacker can influence, including retrieved documents, inbound email, tickets, web pages and tool results.
- **The ability to communicate externally.** The agent can send something out.

Any agent with all three can be turned into an exfiltration tool by a single injected instruction. The hostile text steers the agent, the agent retrieves the sensitive data, and the agent sends it out. No firewall alert fires, because every step was a permitted action performed by an authorized identity.

## The finding that reframes the problem

Here is the observation worth building a control model around, and it comes out of the incident analysis rather than from theory.

In February 2026 a consumer-style agent, asked to review an inbox and suggest what to archive, deleted messages instead. There was no attacker. The agent simply had live delete permissions and weak confirmation controls.

The analysis of that incident makes the point precisely: the permission model behind the unprovoked failure is the same permission model an attacker would exploit through prompt injection.

That collapses two conversations usually held separately. Safety failures and security failures share a root cause. An agent that deletes your records because it misunderstood the request is running the same architecture as one that deletes them because it was persuaded to. The difference is the presence of an adversary, and the permission model is indifferent to that.

Which produces the useful reframing. You cannot reliably prevent an agent from being persuaded. You can determine what a persuaded agent is capable of doing.

## Blast radius, and the limit of permissions

The consensus primary control is blast radius: assume compromise and limit the damage. OWASP’s guidance is explicit about deny-by-default tool access, non-transferable sessions, and no transitive trust between agents.

The standard implementation is permissions. Scope the agent’s credentials, narrow its tool list, require confirmation for destructive operations, and avoid the pattern OWASP catalogs as ASI03, identity and privilege abuse, where agents simply inherit the full permissions of the user who triggered them.

All of that is correct and worth doing. It also has a ceiling, and the ceiling is worth naming.

**Permissions bound where an agent can act. They do not bound what it does there.**

An agent with write access to your ERP, scoped properly and holding only the permissions it needs, can still post the wrong journal entry, approve the wrong invoice, or update the wrong vendor’s bank details. Every one of those is inside its permissions. The credential model cannot distinguish a correct posting from an incorrect one, because both are the same API call made by the same identity.

So permissions reduce the blast radius from the whole estate to one system. Within that system, the radius is still the full set of things the agent could decide to do.

## Bounding capability rather than access

The further control is to constrain not the systems an agent can reach but the actions it can take.

An agent that can execute arbitrary operations within its permissions has an action space nobody can enumerate. An agent that can only execute approved, versioned procedures has an action space that was written down and reviewed in advance.

That difference changes what an injection can achieve. A successful injection can alter what an agent concludes. It cannot cause the agent to perform an action that was never an approved procedure, because there is no path from a belief to an action that does not pass through the rule.

Three properties follow, and they are testable:

- **The action space is enumerable.** You can list what the system is capable of doing, because it is the set of approved procedures.
- **Every action cites a rule.** An action that cannot name the rule that authorized it does not execute, which means an injected instruction has nowhere to land.
- **Changes are versioned and approved.** Modifying what the system can do is a change control event, not a prompt.

## Where Kognitos fits

Kognitos is not a security product. It does not detect prompt injection, filter inputs, or replace the AI guardrail, DLP and agent-firewall tooling that does that work. Those controls remain necessary, and the vendors building them are doing the input half properly.

What the architecture addresses is the capability half. A language model drafts the procedure and a person approves it. A symbolic executor runs that approved procedure, and every write comes from a versioned rule rather than from the model’s judgment. The model reads and interprets; it does not act.

The security consequence is specific. An injected instruction can change what the interpretation layer believes about a document. It cannot introduce an action that no approved procedure contains, because the executor has no mechanism for taking one. The blast radius of a successful injection is bounded by the procedure set rather than by the credential set.

Every run is also recorded in plain English with the inputs, the rule applied and the action taken, which matters after an incident as much as during one.

For related material, see our guides on [shadow AI](/blog/shadow-ai/), [AI agent governance](/blog/ai-agent-governance/), [AI TRiSM](/blog/ai-trism/), [decision lineage](/blog/decision-lineage/), [agent observability](/blog/agent-observability/), and [the agentic AI RFP template](/blog/agentic-ai-rfp-template-2026-vendor-questions/). To see an architecture where the action space is approved in advance, [book a demo](/book-a-demo/) or [try the platform](https://app.us-1.kognitos.com/).

## Getting started

Two exercises, in this order.

**Check your agents against the lethal trifecta.** For each agent in production, establish whether it has access to private data, exposure to untrusted content, and the ability to communicate externally. Any agent with all three is one injection away from being an exfiltration tool, and the usual answer is to design out or gate the third leg rather than to filter the second harder.

**Then ask what your most privileged agent could do if it were simply wrong.** Not compromised, wrong. If the answer is materially the same as what it could do if compromised, your security model and your safety model are the same model, and improving one improves both.

## Frequently Asked Questions

### What is agentic AI security?

Agentic AI security is the practice of protecting systems where an AI agent takes actions across tools and systems rather than producing text. The threat model differs from generative AI because a manipulated agent takes a wrong action in a real system with real credentials, often without a human in the path.

### What is prompt injection and why is it hard to fix?

Prompt injection is the insertion of instructions into content an agent processes, causing it to follow the attacker’s instructions instead of the operator’s. It is hard to fix because a language model receives instructions and data through the same channel, with no syntactic boundary between them. Meta describes it as a fundamental, unsolved weakness in all LLMs, and OWASP maps it to six of the ten categories in its Top 10 for Agentic Applications.

### What is the lethal trifecta?

A formulation attributed to researcher Simon Willison describing three properties that together create structural risk: access to private data, exposure to untrusted content, and the ability to communicate externally. An agent with all three can be converted into an exfiltration tool by a single injected instruction, with every step appearing as a permitted action by an authorized identity.

### What is blast radius in agentic AI security?

Blast radius is the extent of damage a compromised agent can cause. Because prompt injection cannot be reliably prevented, limiting blast radius is the consensus primary control, implemented through deny-by-default tool access, scoped and non-transferable sessions, confirmation gates on destructive actions, and no transitive trust between agents.

### Are permissions enough to secure an AI agent?

No. Permissions bound where an agent can act but not what it does there. A correctly scoped agent with write access to one system can still take a wrong action inside that system, because a correct and an incorrect operation are the same API call by the same identity. Constraining the set of possible actions is a separate control from constraining system access.

### How are AI agent safety failures and security failures related?

They share a root cause. Analysis of a 2026 incident where an agent deleted inbox messages without any attacker involved concluded that the permission model behind the unprovoked failure was the same permission model an attacker would exploit through prompt injection. An agent that acts wrongly through misunderstanding and one that acts wrongly through manipulation are running the same architecture.
