Thought Behind Things
One person with AI may soon do the work of ten
Faizan Khan, founder of OPAG and Senior AI Architect at Secure Agentics, explains why AI agent security is lagging behind adoption, how prompt injection and runtime policy layers work, and what the shift means for jobs.
Contents
Why this conversation matters now
The episode opens with a concern that has been building for some time: AI agents have become capable enough that their security can no longer be an afterthought. Muzamil Hasan frames the conversation around recent security incidents, including a widely discussed case involving an OpenAI agent and Hugging Face, a major platform where open-source models are hosted. He also points to a larger attack on the internet that led many platforms to rotate their keys.
Muzamil says he wants to understand what is actually happening in agentic AI security, how seriously people are taking it, and at what level the work is being done. To answer that, he brings in Muhammad Faizan Khan, founder of OPAG, which builds AI agents that automate repetitive business operations, and a Senior AI Architect at Secure Agentics, a UK-based consulting company. Faizan’s core focus, as Muzamil notes, is cybersecurity, especially agentic security.
The attacks are old, but the scale has changed
Faizan is careful to say that the underlying techniques are not new. “These are old things,” he says in substance. What has changed is that AI is now clever enough to find weaknesses in innovative ways, and when thousands of agents come together and launch brute-force attacks, defending against that volume becomes very hard.
Muzamil pushes him on whether he really believes the situation is serious. Faizan answers with a personal perspective, half-joking that he is “a bit of a conspiracy theorist,” but says he does believe the world is at the edge of a major cyberattack. He compares the possibility to 2020, when physical lockdowns changed daily life, and suggests a digital lockdown could be next. He says people he met in San Francisco, where he recently visited, are sounding serious alarms and admit the industry went deep into AI without doing enough work on its security.
What governance and observability actually mean
Muzamil asks Faizan to explain two terms he used repeatedly: governance and observability. Faizan says governance exists so that an AI agent follows compliance requirements and stays aligned with policies. “That is the core backbone of any enterprise organisation,” he says. He gives a healthcare example: if an AI agent involved in patient care leaks a patient’s personal information, the organisation is no longer compliant.
Observability, he explains, is the tool that lets people monitor and track what an agent is trying to do and the actions it takes. He notes that several larger startups and companies are working on three main pillars: security, governance, and observability.
Faizan then separates two kinds of security concern. The first is whether an agentic system is itself secure, so it does not go rogue, harm customers, or get broken into by outside attackers. The second is the broader exposure of anyone who uses digital tools at all. Muzamil observes that many people in Pakistan have felt safe because they did not touch AI, but Faizan argues that attackers who are themselves using AI can operate at a scale that makes that comfort misplaced.
How runtime security works
The most concrete part of the conversation comes when Muzamil asks how Faizan would approach securing a set of AI systems. Faizan describes his company’s focus as runtime agent security. Most developers, he says, are almost entirely focused on building the AI, and “very less than one percent” are working on making that build secure.
He walks through prompt injection as a central example. An agent’s job is to take whatever message arrives, process it, understand it, and reply. A malicious actor can send a message that is really an instruction in disguise. If the agent has access to contacts or other actions, the attacker can use it to extract information or perform tasks the owner never intended. Faizan notes that OpenAI itself said in 2025 that its latest models would not be fully secure against prompt injection, because the risk comes from how language models work.
Runtime security, as Faizan describes it, sits between the agent’s reasoning and its actions. It tracks the agent’s intent and compares each action against the policies the owner has defined. If an agent tries to share data it should not share, the runtime policy can stop the action before it happens.
Muzamil summarises the approach as a non-deterministic, LLM-based decision layer paired with a deterministic, hard-coded policy layer that judges outputs and rejects anything outside the defined parameters. Faizan agrees, and adds that a runtime policy can be set to block, to alert, or to require human approval. In block mode, the action is stopped. In alert mode, the human is notified and decides whether the agent proceeds. Faizan says the platform he uses is open source and free to self-host, with a cloud version that offers a free tier with usage limits and prebuilt policies for different kinds of agents, such as customer support or financial agents.
Building and securing his own system
Faizan shares how he has built his own workflow. When people apply to appear on his podcast through an online form, AI agents research them, store their profiles, shortlist candidates, extract their information, and send invitations. Once a session is scheduled, another agent starts a conversation on WhatsApp, and in the days that follow, several agents handle the podcast’s management, YouTube descriptions, chapters, and comment responses.
He admits that each agent has its own personality and tool access, and that there are multiple places where the public could inject instructions, including YouTube comments. He says he has tried to solve the problems he can with a straightforward approach, and he is still curious about what other vulnerabilities he should watch for.
Faizan says the main vulnerabilities usually sit in tool access and in the agent’s thinking process. Prompt injection works by using urgency, or by framing a harmful request so that the agent believes it is a legitimate task. He also describes how a runtime security tool can be added to no-code workflows as well as coded ones, so that it sits between what the agent intends to do and the action it takes.
Muzamil asks how much this costs, especially for a small business in Pakistan. Faizan explains that the cloud version has a free tier with rate and usage limits, and that the open-source version can be self-hosted on a virtual machine. He stresses that anyone working on agentic workflows should use some form of runtime security, because the risk falls not only on the individual but also on the organisation and the personal data it holds.
On jobs, creativity, and the human in the loop
Muzamil brings up one of the sharpest questions of the episode: whether AI will replace people. He describes his own experience of moving from employing eight to ten people for a workload that now sits largely with one editor, with the total cost of the remaining work falling to a fraction of what he used to pay. He asks what the people who lost their roles should do.
Faizan’s answer is nuanced. Automation is only part of the picture, he says, and the roles most exposed are those that involve repeatable, process-driven work with little creative input. He points to design work as an example. Designers who relied on searching for inspiration and producing routine logos are more vulnerable than those who understand brand presence and can produce original ideas. He also argues that many people in software were already replaceable if their work was limited to writing code without understanding the architecture or the business impact.
He says that the number of routine roles at risk is significant, but that AI is still not able to fully replace people who combine expertise, judgment, and accountability. He insists that the human in the loop should remain. In his view, you cannot remove every person from the process and grant full autonomy until your security, guardrails, and policies are properly aligned across the whole business and workflow.
Muzamil asks Faizan to look ahead. Faizan expects the next frontier to be physical AI, where systems such as humanoid robots, drones, and power systems interact with the physical world. He says these systems will develop their own ways of planning, deciding, and acting, which makes security even more important as they enter daily life.
Closing thoughts
The episode ends with Muzamil thanking Faizan and inviting viewers to share how seriously they take AI security in their own work, and whether they simply ask an assistant to make their system secure or look into best practices and tools themselves. The conversation leaves a clear picture: agent adoption is moving quickly, security is trailing behind it, and the people who build these systems are the ones who will need to close the gap.
Build with me in the Inner Circle.
You don't need to code. You need a problem you know well, and a room of people building beside you.

