Explore every episode of the podcast Insecure Agents

Most people that architect an AI agent think to ask a binary question: is this tool call allowed, yes or no? Michael Davis, Global Chief Security Architect at J.P. Morgan, argues it's never that simple.

He walks through the four dimensions robotics uses to decide whether it is safe for an arm to move (the state of the system, the state of the environment, the quality of the decision-making, and the uncertainty of the action itself) and maps each one to agents: did the agent gain tools or network connections it did not have before, is it still moving semantically toward its goal after 30 tool calls, and what has the agent already tried.

We get into why  if you think you need memory, you're probably not thinking of your problem the right way. We also explore software factories and why they need to operate as a  cumulative progressive process versus a one shot of go build me this SaaS. Over the course of the episode we piece together what a good reference architecture for agents could look like.

Jet Anderson, is a Distinguished Engineer at GEICO and leads the transformation of their product security function. He joins us to discuss the agentic SDLC and why most of what security teams do has to change while the first principles should stay the same.

We get into why static analysis and human triage no longer keep pace when models write the code, why we need new AI infrastructure, and why the Hugging Face sandbox escape was a decade-old Kubernetes misconfiguration found at machine speed.

Jet tell us the story of when he asked his own agent to "wrap this up" and it merged the PR and did an unauthorized production deploy. He walks through the pre-commit hooks and branch protections he built to prevent this from happening again so that deploys still get an approval gate and an audit trail.

He also explains why teams that skip ideation, design, and specification to go straight to code end up with less secure software the more they iterate, and why the unglamorous answer is doubling down on least privilege, egress control, and build containment.

Travis McPeak, security lead at Cursor, joined us at Black Hat to talk about what causes coding agents to become problematic and what he's done to keep them on track.

What he sees most is the main agent trying hard to solve your problem, getting stuck, and then getting creative. He calls it thrash, and most of the time when somebody says an agent deleted their database, that is what happened.

This episode gets into why static controls like block lists and WAFs break down against an agent whose whole job is to find a way through, why an agent can't monitor itself, and how Cursor's auto review agent works: every tool call the main agent is about to make goes to a separate reviewer that does not care about the main goal at all and only enforces its rules.

Travis also explains why the team had to carefully word the feedback so the main agent would not spiral, why blocking an action can cause the exact thrash you were trying to prevent, and why least privilege finally scales now thanks to new solutions.

Alissa Abdullah, PhD ("Dr. Jay"), Deputy Chief Security Officer at Mastercard, and Arjun Ramakrishnan, Senior Principal Cybersecurity Architect at Mastercard, join us live from Black Hat to explain why access models built for humans start to break when agents act on their behalf.You can't provision access for an agent the same way you provision it for an employee, and just-in-time authorization takes on a new meaning when the consumer operates at machine speed. Dr. Jay and Arjun explain why every tool call and action needs to be authorized, why that's difficult to enforce, and what security teams need when an agent goes rogue, including the ability to revoke access instantly through a button or API call.We also get into Mastercard's three-layer framing of security for AI, from AI, and with AI, and why agent governance needs to move from periodic review to real-time enforcement.

Allie Howe sits down at Black Hat with Aaron Stanley, former CISO at dbt Labs, and Ahmad Nassri, CTO at Socket, to sketch the first draft of a reference architecture for securing software factories.

In the last year there's been considerable advancements that makes now the largest inflection point for software factories. Model vision has improved allowing them to see and verify work they couldn't before. Agents now have access to far richer tool ecosystems and live data, enabling them to work across real production environments. Context windows have gotten larger and reasoning models have improved helping the model think through more sophisticated tasks.

All of those advancements come together now. The race to build a software factory is on, and teams are struggling with how to harness the power of these models while retaining control.

We explore the missing security model for software factories and what the components of a secure reference architecture could look like. We get into the Andon cord problem (can an agent recognize it has been blocked and stop, instead of innovating around the constraint), why the factory cannot have a human's identity and needs purposeful authentication and authorization of its own, why the enforcement boundary has to sit outside the agent loop rather than inside it, and why a single poisoned dependency in a factory is an incident in every work tree at machine speed. Aaron and Ahmad dive deep into supply chain, agent identity, verification, and how to handle what Aaron calls the pernicious problem: a goal seeking agent circumventing constraints to accomplish a task.

An autonomous agent spent days inside Hugging Face production infrastructure and the headline was that it escaped its sandbox. Luke Hinds, founder of nolabs and creator of Sigstore, frames it differently. The agent had root on the execution environment, and "a sandbox is only as strong as the access that you grant to it."

Luke walks us through what it looks like to create an environment where agents have both security and capability. He tells us you can't just lock an agent in a box. If you want the agent to do real work you'll need to delegate some authority to it and give it access to real tools and data.

Luke goes over the three sandbox categories buyers are choosing between today (serverless execution, host isolation microVMs like Firecracker and gVisor, and the fine-grain capability-based approach he is building), and outlines where each one's threat model starts and stops.

Luke explains what nono could have done to prevent the Hugging Face attack, including gating the C compilers the agent leaned on. His parting advice to us is defense in depth, and never let perfect be the enemy of good.

In a world where agents are chaining vulnerabilities together to escape sandboxes, simply blocking bad packages is not enough. Ahmad Nassri, CTO of Socket and previously CTO of npm, joins us live at Black Hat to explain what happens when you deny a coding agent a package: it becomes a risk if the agent thinks it can help it complete its goal later.

Socket has watched agents blocked from an install go straight to the CDN to pull the tarball directly, or rewrite the registry configuration in the local environment and resolve npm by DNS to fetch it another way. For this reason Socket's answer is not a simple denial. Because the enforcement point sits at the network level, Socket changes what the agent and the package manager see in the first place, masking the bad versions so that, as Ahmad puts it, as far as the agent is concerned those versions do not exist.

We talk through the Hugging Face incident where an OpenAI agent found a zero-day in a package registry proxy, how Socket detects a malicious package within minutes of publication, and why agent security is layered: safe packages, short-lived and task-scoped credentials, and real-time visibility into what the agent actually did.

Catherine Jue, co-founder and CEO of Kernel, joins us to explain why browser agents are not blocked by model capability anymore. They are blocked by identity. Kernel builds open source browser infrastructure for AI agents, which means running Chromium in sandboxed Firecracker VMs at scale and solving the part nobody designed for: an agent acting on behalf of a human, on a login page built 20 years ago for a human. Catherine walks through Kernel's Managed Auth product, an SDK that lets developers collect end user credentials so Kernel handles login and re-authentication and the LLM never touches the secret. We get into what agents inherit today (usually the user's full permissions, with no scopes and no audit trail), how ID-JAG and Enterprise Managed Auth for Claude change the picture for MCP servers and what the equivalent looks like for a browser, why multi-hop delegation chains from agent A to agent B to a Kernel browser session have no one-size-fits-all answer, and how intelligent egress networking inside the sandbox can enforce where an agent is allowed to go at the network layer. She closes on the two protocols the early web never standardized, payments and identity, and why Kernel is pushing on the second one.

"I think agents cannot be trusted." That's what Diptanu Choudhury, founder of Tensorlake, told us at AI Engineer World's Fair. Diptanu has built cluster schedulers at Netflix, HashiCorp, and Facebook, and he says the credential model we created for human-authored software does not work for autonomous agents. He joins us to explain why teams are pulling secrets out of the sandbox, why agents are bringing stateful compute back at scale, and why every write in Tensorlake's file system records which agent made it and what credentials it held.

Software factories are technically possible today, yet almost nobody can is operating one. The security model is what's missing. In one incident a coding agent deleted PocketOS' production database as a side effect of an unrelated fix. In another, a distinguished engineer at GEICO asked an agent to land a pull request and watched it push to production instead. The reality is coding agents cannot yet be trusted with the autonomy a factory requires.

Recorded at The Marquee in Las Vegas on the backdrop of Black Hat, Allie Howe moderates a panel with Eli Aleyner, VP of Product Strategy and Alliances at Docker, Ian Livingstone, CEO and Co-Founder of Keycard, and Ezra Tanzer, AI Forward Deployment CTO at Snyk, on the security problems standing between teams and the autonomy curve a software factory requires.

The three companies co-authored Agent Baseline, a vendor-neutral reference architecture published days before the panel that defines agent security by six outcomes (Discover, Constrain, Authorize, Observe, Validate, Respond) and 35 capabilities rather than by product category. We get into why most teams believe they can only pick two of security, capability, and autonomy, Eli's (Docker) response to the Hugging Face/OpenAI sandbox escape, why credentials belong injected at the moment of use instead of sitting in the agent's workspace, and why customers are seeing risk come from negligence and over-provisioned agents rather than from malicious engineers.

Andrew Baker, who leads Developer Relations at Temporal, and Cornelia Davis, Principal Technologist at Temporal and author of Cloud Native Patterns, join us from AI Engineer World's Fair to explain why building AI agents keeps re-teaching the industry lessons it already learned in the microservices era. We get into how MCP is growing up, moving from a simple request-response protocol to async MCP Tasks, going stateless, and adding an extension for OAuth and Enterprise Managed Auth. Cornelia walks through why shared memory across parallel agents brings back both an access control problem and a concurrency problem the industry spent years designing away, and why it is still very much unsolved. Andrew explains how the harness and the credentials are moving outside the sandbox, so an agent holds only a short-lived token scoped to the task at hand and nothing more. And they show why a durable event history is what lets you reconstruct what an agent did and roll it back, so that if an agent drops a database, you can actually go back.

Sergey Burykin, Senior Software Engineer on Uber's AI Security team, joins us to explain the agent identity crisis and how Uber solved it while running roughly 1,000 agents in production. 

Sergey helped write Uber's article "Solving the Identity Crisis for AI Agents," and his core argument is that an agent should be authorized on the intersection of user permissions and agent permissions, never just one. Use only the user's permissions and a hallucinating agent can make calls the user never intended. Use only the agent's identity and any user who reaches the agent inherits access to sensitive business and customer data. 

We get into the infrastructure Uber built to enforce that (a secure token exchange service and an MCP Gateway as the policy enforcement point), why AI security depends on identity, authorization, runtime guardrails, and observability, and why static OAuth scopes break for non-deterministic agents that need dynamic, least-privilege access.

Manoj Nair, CTO and Chief Innovation Officer at Snyk, joins us at Snyk HQ during AI Engineer World's Fair to discuss the architectural decision he argues the next 24 months of agentic security depend on: the generator cannot be the validator.

We get into why "the fox guarding the henhouse" is suddenly a live security question ("I can use AI to secure AI, so do I still need a separate security offering?"), why finding vulnerabilities was never the hard part and fixing them safely inside the loop is, and why securing at inception means combining independent models with deterministic data and security research rather than just pointing one model at your codebase.

Manoj also walks through Snyk's Evo, an agentic security orchestrator built on the fighter-pilot OODA loop that turns a security engineer into a 10X AI security operator, why the agents' own supply chain (the MCP servers and skills they pull at runtime) is now its own attack surface, and why he sees security as the throttle, not the brake.

Diana Kelley, CISO at Noma, has spent years on the front lines of enterprise security across IBM, Symantec, and Microsoft, and now she is helping write the rulebook for the agent era. She joins us to make the case that the cloud shared responsibility model does not translate to AI. 

In the cloud there were roughly two responsible parties and your data was always your data. With agents there are at least three, the frontier model provider, the platform or developer building on it, and the user, and the trust boundary has moved from storage to decision-making and action. 

We open on the PocketOS incident, where a coding agent used an over-scoped Railway token to delete a production database and its backups in nine seconds, and use it to trace where responsibility actually lives. 

Diana then walks through AARM, the runtime security specification she co-chairs at the Cloud Security Alliance, why authorization needs five decisions instead of two (allow, deny, modify, step-up, and defer), how much context an agent can actually trust, and why the most useful question a CISO can ask a vendor is not "where does your responsibility end" but "can you sit down and explain how you threat modeled this."

AAuth Night: Moving Beyond OAuth was an AI Engineer Side Event during World's Fair on July 1st 2026.

We explored the challenges with agent auth today, the best practices available today, and future solutions such as new protocols like AAuth that are in the works.

OAuth, JWTs, and API keys were built for humans and servers, not agents that act on your behalf, chain tasks across tools, and decide what they need at runtime. As AI engineers are shipping agents into production they begin to feel the problems with agent auth and identity today. Consent fatigue, credential management, and agent alignment start to show up and create friction.

This panel explores the problems with agent auth today, how we are updating OAuth to be better for agents with grants like ID-JAG, and why we ultimately need to build a new protocol for agents (AAuth).

This panel was recorded at AAuth Night: Moving Beyond OAuth, an AI Engineer side event on July 1st, 2026 and brings together four people rebuilding auth for agents: Dick Hardt, founder of AAuth and OAuth, Karl McGuinness, identity expert and past Chief Product Architect at Okta, Ian Livingstone, co-founder and CEO of Keycard, and Herman Errico, founder of AARM.

Guy Podjarny built Snyk into the company that taught developers to secure their dependencies. Now, with Tessl, he argues that agent skills have become a new unit of software, one that deserves the same rigor we give source code. Guy Podjarny, founder of Tessl and Snyk, joins us to explain why skills are the new code: context is the only layer that runs straight inside the model's reasoning loop, so it is effectively the programming language for models. We get into the new supply chain that follows (there are already 2 million skills in the open ecosystem, up from near zero last September), why a malicious or negligent skill is so hard to catch, why there's still no npm or PyPI for skills, and why enforcement is moving into the harness because the model cannot police its own context. Guy also makes the case that skills rot, so a skill you write today can be useless or harmful three months from now without a plan to maintain it.

Every SaaS app an enterprise connects to stands up its own OAuth stack of long-lived grants the enterprise can't see or revoke. Karl McGuinness, author of ID-JAG and past Chief Product Architect at Okta, calls these "OAuth islands," and agents turn them from a nuisance into a serious risk. He joins us to explain OAuth federation, how ID-JAG shipped inside Anthropic's Enterprise Managed Auth for Claude, and what it really takes to govern agent access from one central checkpoint.

This is one of the best public internal AI stories we've seen, built by just a few engineers. Derek Meegan, a software engineer at Browserbase and the lead behind their internal AI agent, bb, joins us to explain how bb took feature-request coverage to 100% with zero human effort, got 99% of support first responses under 24 hours, and turned 30 to 60 minutes of manual log-diving into a single Slack message.

Then we get into the part most teams skip: the security model that lets you actually trust an agent with that much reach.

Derek covers code mode, the sandbox that never touches a secret, credentials brokered just-in-time through an integration proxy, least-privilege tools on every event-driven trajectory, and permissions computed per invocation instead of written into a static config.

His thesis: use the agent to take the repetitive, well-understood work off people's plates, and make the harness verifiably secure, because security is what lets you scale it.

David Cramer, CPO and co-founder of Sentry, joins us to cut through the agent hype with a working engineer's skepticism: the model is rarely what holds agents back. The harness you build around it is. We get into the Railway incident, where a coding agent found a stray CLI token and deleted a production database (and every backup) in nine seconds, and why the enforcement layer has to live below the agent, not in an advisory system prompt.

David explains Seer, Sentry's AI debugger, as the counter-example: an agent doing real work because it was given the right context, not more autonomy. He also walks through Warden, the code-review harness he built that found 100+ previously unknown vulnerabilities across Sentry and open-source projects, including full auth bypasses, for roughly $1K of compute.

We also get his contrarian-but-consistent take on why MCP is not just a shim on your API, why CLIs are harder to secure than people think, and why verification, not code generation, is still the unsolved problem.

Herman Errico, Product Manager for Technical Research at Vanta, joins us to discuss AARM (Autonomous Action Runtime Management), the spec he created to define a brand-new security category for agents that take real actions, not just generate text. We get into why the action boundary is the security boundary, why securing the model, prompt, or orchestration layer is the wrong place to enforce, and why a runtime needs five authorization decisions (allow, deny, modify, step-up, and defer) instead of a binary yes or no. Herman also explains why he didn't ship a product but a spec, then donated it from Vanta to the Cloud Security Alliance so the industry can compete on execution instead of marketing, how to reason about which context an agent can trust, and why you must block the action from occurring before it takes place.

Malte Ubl, CTO at Vercel, joins us to discuss deepsec, Vercel's open-source AI security harness designed to scan entire codebases for vulnerabilities using coding agents like Claude and Codex. We explore why software engineering is shifting from programming models to programming agent harnesses, how deepsec scales security reviews across millions of lines of code, when AI token spend is justified, and why Vercel is betting on AI Gateways, microVM sandboxes, and self-driving infrastructure to power the next generation of software development.

Sunil Agrawal, CISO at Glean and one of the authors of the AWARE Framework, joins us to discuss the new guide for governing generative and agentic AI he co-authored with Palo Alto Networks and Databricks. This framework gives CISOs a much needed playbook in a rapidly evolving threat landscape. Palo Alto's Unit 42 showing AI-assisted attacks can now reach data exfiltration in as little as 25 minutes, leaving defenders almost no time to respond.We dig into AWARE's five behavioral dimensions, why governing modern AI means controlling intent and context rather than just access, how to give every agent a scoped identity instead of shared credentials, and the cascading risks that emerge when agents start delegating to other agents.

Alex Stamos, CPO of Corridor and past CISO at Facebook, and Andrew Becherer CISO at Socket, join us to discuss the open letter they and 100 others have signed in opposition to the US government taking down Fable after research from Amazon showed capabilities that gave the current administration pause.

We discuss the potentially dangerous precent this sets, the state of the letter, and what to do while waiting for Fable to come back online.

Why does every AI security incident seem to trace back to auth? We sit down with Damian Schenkelman, VP of Research and Development at Auth0 to discuss⁠ recent incidents in the news, MCP, the act claim chain, and the future of agent identity.

The conversation digs into the core problem agents create: when an agent hands a task to a sub-agent, which calls an MCP server, which hits a SaaS API, who is actually making this call, and on whose behalf?

In this episode we sit down with Dick Hardt, the creator of OAuth, to talk about why the auth primitives we built for the web fall apart the moment agents start acting on our behalf. We dive in to why OAuth doesn't fit MCP, what breaks when an agent runs for hours and touches a dozen systems using your credentials, and his new protocol, AAuth: a way for developers to run agents without API keys.

We sit down with Geoff Huntley, creator of the Ralph Wiggum Loop and founder of LatentPatterns.com, to hear his take on where AI is pushing software next: hyper-personalized software, software factories, and eventually product factories that optimize themselves for revenue.

With this level of hyper-personalization that AI now allows for Geoff says he finds himself asking vendors "are you a utility company, or are you something I'm going to clone?" He walks us through what this means for moats, open source, and the future of how software gets built.

We sit down with top AI engineers such as Sherwood Callaway, founder of Sazabi, Anthony Shew, core maintainer of turborepo at Vercel, and Dexter Horthy, CEO of HumanLayer, to hear about how they are using sandboxes to make agents more performant. We also discuss the security differences amongst sandbox providers with Rene Brandel, founder of Casco. We discuss how sandboxes aren't created equal and how the underlying infrastructure that powers them directly impacts your security posture.

Mark Dorsi, CISO at Netlify, sits down with us at RSAC to talk about the shift to everyone becoming a builder and how he's coding 6 hours a day and how products, including Netlify, must adapt to a world where most users are agents.

Kyle Bhiro and Josh Kotrous from Pensar join us at RSAC to discuss how AI is reshaping the entire AppSec industry. Kyle and Josh elaborate on how agentic code scanning and continuous testing is leading to AppSec market consolidation and new expectations around AppSec spend.

We also explore the thought that point in time audits may make less sense for AI, which changes constantly. Given this, continuous testing seems to be an emerging standard.

Ian Webster, CEO and Co-Founder of promptfoo, joins us at RSAC to discuss OpenAI's recent acquisition of promptfoo. Ian discusses how appealing to both developers and security teams was key to promptfoo's go-market-strategy strategy.

Ian's success offers a playbook for other AI security companies that may be targeting an acquisition and shares what's next for promptfoo at OpenAI.

Alex Stamos, former CISO of Facebook and current Chief Product Officer at Corridor, explains how AI is reshaping the kill chain and enabling new capabilities for attackers worldwide.

He also outlines what’s needed to defend against these emerging threats and how to prepare your organization for what’s coming.

Animesh is the CEO and founder of PlayerZero, a company using context graphs to build a complete picture of how your production software actually behaves.

Animesh's X article on context graphs went viral getting over 2M views. Animesh explains what a context graph is, why you should build one, and how it can help you build a world model around why decisions get made.

The agent identity conversation is back on the Insecure Agents podcast.

Developers are starting to feel the pain of missing agent identity infrastructure as they think through problems like agent memory access and storage and goal based authorization for tools and resources unplanned for at agent inception. Listen to Pavan and Aaron explain how their recent Fine-Grained Authorization (FGA) launch can help.

James sits down to tell us about OpenClaw using Convex, how proper architectural building blocks sets you up for better security, and how the shift to agents writing all of software changes who platforms like Convex are building for.

You've heard of OpenClaw, but have you heard of Momo?

Momo is built by Cailyn Yong and is a personal assistant agent for teams. Momo's memory actually works and makes it stand out against other agents such as OpenClaw. Hear from Cailyn on the AI security issues this type of agent faces, how she built Momo, and what it takes to be successful in this increasingly popular space.

In this episode we discuss the engineering and security challenges that separate POC agents from enterprise agents.

Kwindla brings a wealth of knowledge on common hard agent engineering problems such as async, automatic, non-blocking context compaction, agent memory, and stateful long running agents.

Listen in to learn how Peter created the best personal assistant agent to date and the security concerns at play. Personal assistant agents need lots of access to do meaningful work but there are tradeoffs between innovation and security.

Supply chain security for open source dependencies, how to protect yourself against attacks like Shai Hulud 2.0, and how AI agents introduce new security challenges.

Agents need purpose-built sandboxes that spin up in milliseconds to execute tasks like code analysis, web browsing, and data processing. Ivan addresses the hurdles around speed, security, and statefulness.

An attacker can hide prompt injections in images that only become to AI systems, enabling data exfiltration on production systems like Google Gemini CLI. Is weaponized image scaling a security vulnerability, or an architectural flaw in how AI systems process multi-modal inputs?

We sat down to discuss the just released OWASP Top 10 for Agentic Applications, exploring critical threats like goal hijacking, remote code execution, and identity management while breaking down how to balance AI agent autonomy with deterministic guardrails and user trust.

Browser agents need standardized ways to identify themselves and prove their legitimacy when accessing the web. We take a deeper look at credential management, scoped permissions models, telemetry for monitoring behavior, and implementing hard boundaries to prevent prompt injection and unauthorized actions for browser agents.

The highly anticipated MCP debate. We explore critical questions around SDK replacement, marketplace curation, enterprise concerns, authentication challenges, and whether MCP represents a security nightmare or the future of agent systems.

Bryan and Alex discuss how AI agent architecture directly impacts security posture. We take a look at everything from infrastructure-level guardrails rather than relying solely on tool-call layer protections to the cold start problem and defense-in-depth strategies against prompt injection.

Samuel Colvin founded Pydantic in 2017 and launched the company in 2023. He discusses MCP security vulnerabilities, AI agent authentication challenges, and the upcoming Pydantic AI Gateway for threat detection.

Mackenzie joins us to discuss AI in code security, smarter vulnerability prioritization, and Aikido's research into malicious packages in open source.

Steve Vandenburg, AI Security Architect at Cotiviti, discusses the evolving role of AI security in enterprise environments and how frameworks like NIST AI RMF, HITRUST, and the new SAIL framework translate from policy into real technical implementation.

Dor Sarig has spent nearly two decades in cybersecurity, from offensive work with the Israeli government to leading product roles at Simulate and Perimeter 81. Now CEO of Pillar Security, a unified platform to secure the entire AI lifecycle and is behind the SAIL framework.

This week we're taking a deep dive on the agent identity problem. Ian Livingstone, Matt Creager and Jared Hanson founded Keycard to accelerate agent adoption without sacrificing control.