


The episode explains “authenticated delegation” for AI agents—how to prove an agent’s identity, that it acts on behalf of a specific user, and that it has only task-specific permissions (not carte blanche). It argues this is needed because AI agents have broad, high-risk capabilities (web/API access, multi-agent chats) and because natural-language instructions are brittle (prompt injection).
No guests are named in the transcript; it’s presented as a host-led discussion of a paper.
Use token-based delegation (extending OAuth 2.0 and OpenID Connect) with three tokens: user ID token, agent ID token (capabilities/limits), and a delegation token signed by the human specifying scope, validity, and spending limits. Enforce structured, machine-readable resource scoping (firewall-like rules) plus human review. This improves privacy/contextual integrity and reduces legal ambiguity via explicit authority.
Holiday booking—agent should use trusted payment services with enforced spending caps, not raw credit-card handling. Telemedicine—text-only agent must be flagged so portals require human review for multimedia data. Multi-agent insurance claims—prevents hospital-agent impersonation and fraud. Air Canada bereavement-fare chatbot—company held liable due to apparent authority.
Written by AI. May contain mistakes.
Explore what authenticated delegation means and its significance in AI.
“technical maybe a little academic we've all seen basic app permissions right allow this app to access your camera click okay but you're saying with ai agents this whole concept gets way more complex.”
Discuss real-world scenarios of AI agents booking travel and associated risks.
“The paper has some great examples that really drive home why this isn't just theoretical.”
Examine the consequences of AI agents having limitations in processing information.
“Okay, but beyond just money, what about situations where an agent has limitations?”
Understand the challenges of trust and authentication between multiple AI agents.
“it really highlights how important it is to communicate not just what an agent can do, but also what it can't.”
Learn about contextual integrity and maintaining privacy through authenticated delegation.
“They'd know the agent is acting transparently on behalf of a specific verified human user.”
Delve into how to build secure systems for authenticated delegation using existing frameworks.
“It ensures the assistant sticks to the right context and rejects actions that try to cross boundaries.”
Discover the importance of structured permissions in managing AI agent actions.
“Now, the paper also mentions something called W3C verifiable credentials, or VCs.”
Learn why human approval is crucial in AI agent access control.
“Maybe you need to correct it or refine it.”
Discover the complexities of legal responsibility when AI agents err.
“What happens if, despite all these safeguards, something still goes wrong?”
Understand the critical role of authenticated delegation in AI safety.
“And it aligns well with other legal trends, too.”

Get the gist, explore an idea or ask about something said.
On Best AI papers explained, what are the key points of this episode?
Automatic transcript. May contain errors.
0:00Imagine your AI assistant seamlessly booking travel, managing finances, and negotiating deals. for you. These autonomous AI agents are becoming really powerful, like extensions of our digital selves. But, you know, as they take on more complex tasks for us, a pretty critical question pops up pretty fast. How do we make sure they're doing exactly what we asked? And importantly, nothing more. Today, we're doing a deep dive into a really interesting paper that tackles this exact challenge. It's called AI agents need authenticated delegation. And this isn't just theory, right? it's about building the secure foundation the rails for a future where ai agents are truly reliable trustworthy digital delegates acting on your behalf our mission today is to unpack this it's a complex idea but we want to show you why it's absolutely vital for unlocking the power of ai agents safely mitigating those practical risks okay so authenticated delegation sounds a bit technical maybe a little academic we've all seen basic app permissions right allow this app to access your camera click okay but you're saying with ai agents this whole concept gets way more complex.
1:01What's fundamentally different here? Why the urgency? It absolutely does. A whole different ballgame. At its core, authenticated delegation is really a way for a third party, could be a website, another AI, even a person, to check three key things when an AI agent interacts with them. Is this really an AI agent I'm talking to? Is it acting on behalf of a specific human like you? Right. Linkage to the user. And third, and this is crucial, Has it been given only the specific permissions it needs for this task? Not carte blanche. So, yeah, it's about establishing identity, defining that scope very clearly.
1:38What's allowed, what's not, and making sure the whole thing's auditable. You can track it. I can see the parallels with existing stuff like app permissions. But, yeah, the autonomous nature of these agents. It feels like it just raises the stakes massively. They can browse the web. They can talk to APIs. You know, there's digital handshakes between software. They can even chat with other agents using natural language. How does that huge action space create unique risks compared to, say, my boring calendar app? Well, that's exactly it, your calendar app. It has a pretty narrow, defined set of things it can do.
2:09Check dates, ad events, maybe reminders. AI agents, though, by their very design, they're meant to be general purpose. Highly capable, almost like a digital intern, but potentially way more powerful. And that broad action space means the potential for things going wrong, unintended consequences, or even malicious misuse just explodes, which makes having explicit, verifiable delegation not just nice to have but essential, absolutely crucial. Okay, let's make this concrete. Let's pivot to some real-world scenarios. The paper has some great examples that really drive home why this isn't just theoretical.
2:45Can you walk us through the securing tool use case? Like let's say my AI agent is trying to book me a holiday. Right, good example. So the agent searching the web for flight deals or hotel recommendations, probably fine, doesn't likely need special authorization for that. But what about accessing your personal calendar to check if you're free or even more critically, actually making the purchase, booking the flight, paying for the hotel? Now, you could maybe just paste your credit card number into the agent's chat window. Oh, I would never do that. That sounds terrifying. Exactly. It's incredibly risky.
3:20You're relying entirely on the AI's internal programming, its reliability, its trustworthiness not to mess up or worse, get hacked or tricked. So the real insight here, the paper argues, is to leverage the systems we already have, the secure payment systems. The agent should be authenticated, proven to be your agent, and authorized given permission to make purchases only on specific services, say the airline's website, where your credit card details are already stored securely. Ah, okay. And crucially, with explicit spending limits enforced by that delegation, not just relying on the AI to be good.
3:54That's a huge aha moment for me. It's not about the AI handling the raw credit card data itself. It's about giving it permission to use the existing trusted financial pipes, but with strict rules. Augmenting security, not trying to replace it in a risky way. Precisely. This is a critical shift. Okay, but beyond just money, what about situations where an agent has limitations? Maybe it can't understand certain types of information and the service it's talking to doesn't know that. That's another really profound point. And the paper gives a pretty stark example. Imagine an AI agent used by a doctor in a telemedicine portal.
4:30The agent logs in using the doctor's basic credentials. The portal thinks, OK, this is the doctor or at least has the doctor's full authority. So it grants access to all patient records. OK. Including, say, a video consultation file, maybe with a voice recording from a specialist discussing nuances. But here's the catch. This particular AI agent, it's text only. It can't process video or audio. Oh, no. So it generates its summary or recommendation based only on the text notes it finds. It completely misses the crucial information in the video and audio. The doctor, maybe busy, trusts the AI's summary, assuming it saw everything, and potentially makes a misinformed treatment decision based on incomplete data.
5:10But if that agent's text-only limitation had been explicitly communicated through authenticated delegation, like a flag in its agent ID token, the portal could have reacted differently. It could have said, hey, there's multimedia content here. This agent can't process it. Human review required. Avoiding a potentially harmful oversight. The implications there are huge. It's not just financial risk. It's professional trust, patient safety. it really highlights how important it is to communicate not just what an agent can do, but also what it can't. Exactly. Capabilities and limitations. Both need to be clear.
5:45Okay, so what other surprising areas does this authenticated delegation touch? Well, if we zoom out a bit, think about multiple AIs talking to each other. Multi-agent communication. Let's say you have a hospital's AI agent needing to coordinate with an insurance company's AI agent to process a claim. Right. Happens all the time with humans. Yeah, but now with agents, usual authentication is paramount. How does the insurance agent know it's really the hospital agent and not some scammer? Without authenticated delegation, a malicious third-party agent could maybe impersonate the hospital and submit fake claims.
6:19Oh, wow. Fraud on a whole new level. Or conversely, the insurance agent might just reject valid claims because it can't be sure they're authentic. It introduces friction and mistrust into these automated processes. And sticking with that theme of trust online, AI is getting scarily good at mimicking humans. How do we protect online spaces? Social media, forums, whatever. How do we know who's real and who's a bot? Exactly. This is a huge challenge. Authenticated delegation offers a more nuanced solution than just trying to block all bots, which is often a losing battle anyway. Instead, platforms could explicitly allow certain AI agents, ones that are properly authenticated and delegated by a verified human, to access services in controlled ways.
7:01So not just block, but allow with transparency. Precisely. They'd know the agent is acting transparently on behalf of a specific verified human user. For example, your agent might access your social media with permission to help you draft replies or schedule posts, but it does so accountably. It's a much smarter approach than the old robots.txt file, which is really just a suggestion that polite bots follow. This provides actual verification. Okay, and how does all this tie back to our own digital privacy, making sure our data isn't misused across different contexts? It connects directly to this idea of contextual integrity, keeping information appropriate to its context.
7:40An AI assistant using authenticated delegation could have different sets of credentials, different digital IDs for different roles, Say one identity for when it's helping you with work tasks using your enterprise login and a completely separate identity for when it's helping you manage personal things like booking that holiday we talked about. So it can't mix the two. Exactly. The system enforces these scope credentials. It ensures the assistant sticks to the right context and rejects actions that try to cross boundaries. Like it couldn't or shouldn't be allowed to pull information from your confidential work documents to help fill out your personal tax forms.
8:16Right. It helps maintain those boundaries, those privacy norms that are really important in our lives, but can get blurry fast with powerful AI. Okay, so we have established the why. The need is pretty clear. Security, safety, trust, privacy. Let's get into the how. How do we actually build this? Are we inventing completely new technology from scratch? Great question. And the good news is, mostly no. The paper argues strongly for leveraging stuff we already use, things that are battle-tested on the Internet, specifically OpenID Connect and OOth 2.0. Okay, those names ring a bell. They're used for logins and permissions everywhere, right?
8:52Exactly. They are the workhorses behind things like login with Google or letting that photo editor app access your cloud photos. Ophioth 2.0 is the framework for delegated authorization. You granting one service permission to access resources held by another service on your behalf. And OpenID Connect, or OIDC, is built on top of OFUS 2.0. It specifically handles the user authentication part, proving who you are, and provides what's called an ID token, containing verified information about you. Right. So the core ID in the paper is to reuse and cleverly extend these existing, well-understood patterns to create a new, secure way for humans like you to delegate specific tasks with specific limits to your AI agents.
9:36And the key innovation they propose is a new kind of digital token, like a special ID card just for the delegation itself. That's a great way to put it, yes. The framework relies on three main types of tokens working together. First, you have your standard user ID token. That's you, the human. Your verified digital identity from something like Google or your company login. Okay, got that. Second, there's an agent ID token. This is issued for the AI agent itself. It's like the agent's own ID. It contains its unique identifier, maybe information about its capabilities like text only, or can access financial APIs and potentially built-in limitations like that spending cap we mentioned.
10:10Agent ID, check. And third, the crucial new piece, the delegation token. This is the linchpin. This token explicitly authorizes that specific agent, identified by its agent ID token, to act on behalf of you, identified by your user ID token. And critically, this delegation token is issued and digitally signed by you, the human delegator. It acts as your explicit instruction. So I create that token effectively. Or more likely, you authorize its creation through a user-friendly interface, but your digital signature is on it. And it references both your ID and the agent's ID. It also includes the specifics.
10:47What's the scope of the delegation? Which actions are allowed? How long is it valid for? Maybe even where logs should be sent. This creates a verifiable chain of authority. User delegates to agent. It's cryptographic, really hard to fake. That makes so much sense. It's like giving your AI agent that tailored, verifiable digital permission slip signed by you with really clear instructions and limits. Now, the paper also mentions something called W3C verifiable credentials, or VCs. Are those an alternative to this OIDC-OF approach, or do they fit in somehow? They can be either, really, or often a complement.
11:22VCs are another emerging standard focused on digital credentials you control yourself. A key advantage of VCs is selective disclosure. You can prove just this specific piece of information needed like I am over 18 without revealing your full identity or date of birth. That's great for privacy. Oh, interesting. They're also typically not tied to one big identity provider like Google, more decentralized. So while OIDC and OIF have this huge existing ecosystem, making them maybe easier to adopt now, VCs offer powerful features, especially around privacy and decentralization. The paper suggests that maybe the most practical way forward involves hybrid approaches.
11:56For instance, maybe that agent ID token could contain or embed a verifiable credential that specifies its capabilities. You get the best of both worlds. OK, so we have ways to authenticate the agent and delegate authority using these tokens, maybe even combining standards. But here's a big question. How do we make sure the agent actually sticks to the permissions we've given it? We've all seen AI do weird things or misinterpret prompts. Relying just on please don't access my financial files in the chat seems flimsy. It is flimsy. You're right. That's absolutely critical. Relying only on natural language prompts for security is inherently brittle.
12:31It's vulnerable to prompt injection attacks where someone tricks the AI through clever wording and just plain old misinterpretation by the AI. Language is ambiguous. So what's the robust solution? The most reliable strategy, the paper argues, is to enforce resource scoping with structured permissions. Structured permissions. What does that mean exactly? It means defining the rules in a clear, unambiguous, machine-readable format, not just natural language. These rules specify exactly which resources, like which specific files, database tables, or API endpoints the agent is allowed to interact with and how read, write, delete, etc.
13:08Think of it like firewall rules, but for AI agents. They act as a hard safety net, one that's largely immune to the fuzziness of language or weird behaviors of the AI model itself. Okay, so if I tell my agent, hey, organize my files for Project Alpha, you can read and write in those folders, but definitely do not touch anything in my Finance 2023 folder, that natural language instruction isn't the actual control. Exactly. That's your intent. It's the starting point. But the system, perhaps using another AI, needs to translate that instruction into a formal structured policy. So it would generate a rule that explicitly lists, say, Files Project Alpha as allowed for read-write access and explicitly denies any access to Files Finance 2023.
13:48I see. It makes expertise. Yes. And here's the crucial final step. You, the human delegator, need to review that generated structured policy. Check that the translation correctly captured your intent. Maybe you need to correct it or refine it. And then you approve and finalize it. Ah, the human in the loop part. Absolutely. This human review and approval of the structured access controls is vital. It gives you the ultimate oversight and ensures the unambiguous rules match what you actually wanted before the agent gets the keys. Right. So we've got authentication, delegation via tokens, and structure permissions reviewed by a human.
14:25That sounds pretty comprehensive. But let's talk about the inevitable. What happens if, despite all these safeguards, something still goes wrong? An agent makes a mistake, maybe finds a loophole, maybe causes some harm. Who's legally responsible? Who's on the hook? Yeah, this is where it gets really interesting and intersects with longstanding legal principles. We enter the realm of the law of agency. This is a body of law dealing with situations where one person, the principal, that's you, the user, authorizes another person. The agent, traditionally human, now potentially AI, to act on their behalf.
14:57Okay, agency law. A really important concept within agency law is apparent authority. This means a principal can sometimes be held responsible for actions taken by their agent if a third party reasonably believed the agent was authorized to perform those actions, even if the principal never explicitly gave that specific permission. So what it looks like the agent can do matters legally. It can, yes. It's about the perception created. And this is where that Air Canada chatbot case the paper mentions becomes so relevant. Right. Tell us about that one. It sounds fascinating. It really is. So a customer interacted with Air Canada's chatbot and got some incorrect information about bereavement fares.
15:40Air Canada later argued basically, hey, that was just a chatbot. It's separate. We're not liable for its mistake. Interesting defense. Did it work? No, it didn't. The judge essentially ruled that the chatbot operates as part of Air Canada's website, part of its digital presence. Therefore, Air Canada was responsible for the information it provided. The chatbot had apparent authority. Wow, so companies can't just say it was the AI's fault. It seems increasingly unlikely, especially if the AI is presented as acting for the company. This case really underscores that organizations deploying AI agents are likely to be held responsible for their actions.
16:15And this is another reason why authenticated delegation is so important. It makes the scope of authority explicit and verifiable. It reduces the ambiguity and the reliance on fuzzy apparent authority. If there's a signed delegation token clearly stating the agent wasn't authorized for a certain action, it strengthens the principal's position significantly. It provides that clear chain of command, that audit trail. So it brings clarity, which reduces legal ambiguity and hopefully builds trust for everyone involved, the user delegating, the agent acting, and the third party interacting with it. Exactly.
16:48And it aligns well with other legal trends, too. Think about things like the Uniform Electronic Transactions Act, UADA, in the U.S. It talks about the importance of agreed-upon security procedures for electronic records and transactions. If you use a system like authenticated delegation, which is a strong security procedure, and follow it, but an error still occurs, the legal standing might be different than if you had no real procedures in place. It also directly supports that human in the loop principle we discussed, which is increasingly showing up in regulations like the EU AI Act, especially for AI systems deemed high risk.
17:22Having that human oversight baked in is becoming a legal expectation, not just good practice. Okay, wow. We have covered a lot of ground today. From the absolute necessity of securing these powerful AI agents, through the technical nuts and bolts of how protocols like Oath and OpenID Connect can be extended with new tokens, to the crucial role of structured permissions and human oversight, and finally touching on these really deep legal questions of agency and liability, it seems crystal clear that as AI agents get woven deeper into the fabric of our digital lives, Frameworks like authenticated delegation aren't just optional extras.
17:59They're going to be absolutely essential for these agents to operate reliably, securely, and responsibly. I completely agree. And while, you know, no system is perfect, there are always alternative viewpoints, maybe some technical complexities to iron out. And the challenge of translating human intent perfectly into structured rules is always there. The argument for building and implementing robust authenticated delegation systems now is really compelling. We can't let the perfect be the enemy of the good here. We need mechanisms to ensure these powerful AI tools can be safely integrated into everything we do online, building on the trusted infrastructure we already have.
18:33Absolutely. So a final thought to leave everyone with. As you listening start to interact more and more with AI agents, maybe you already are, think about this. How would having a clear, verifiable chain of command, knowing exactly what you've authorized your digital delegate to do, signed off by you, how would that change your trust? How would it change your willingness to rely on these incredibly powerful tools for more important tasks, something to mull over?
This position paper proposes that the widespread deployment of AI agents necessitates authenticated delegation to address challenges in authorization, accountability, and access control. The authors suggest extending existing OAuth 2.0 and OpenID Connect protocols with AI-specific credentials and delegation mechanisms, allowing users to securely grant specific authorities to AI agents. This framework aims to ensure AI agents act within defined permissions and scope limitations, enhancing security and providing clear chains of accountability. The paper explores how natural language permissions can be translated into auditable access control rules, highlighting the legal implications and practical benefits of such a system for managing AI interactions in various digital environments.