Cybersecurity is often framed as a discipline focused on protecting corporations, governments, and critical infrastructure. For many professionals in the industry, the work stops there. Yet these same skills used to defend enterprise systems can also be leveraged to support vulnerable communities, grassroots organizers, and mutual aid networks. This reframes cybersecurity knowledge as something that is not solely confined to corporate settings, but instead applicable to broader and more impactful community efforts. When these skills are shared in service of the greater good, they help bridge the gap between industry professionals and the general public, advancing social justice in an increasingly digital world.

This talk examines how grassroots anti-surveillance efforts, which emerged in response to intensified immigration enforcement operations in Minneapolis, MN, can inspire cybersecurity professionals to apply their expertise toward broader mutual aid initiatives. Using a real-world case study of teaching civilians and organizers the fundamentals of cyber hygiene, mobile security, and government surveillance awareness in a warehouse basement during the ICE occupation this past winter, the presentation demonstrates the meaningful impact that can occur when technical knowledge extends beyond corporate environments and into local communities. By the end of the session, attendees will leave with a deeper understanding of the social implications of digital security, stronger strategies for presenting cybersecurity education in digestible formats, and inspiration to use their technical expertise to support collective care and community resilience. Participants may even leave with a few workshop ideas of their own to bring back to their communities.

This is the 15th BSidesPDX, so we’ll take a walk back into the history behind the conference

Esteban, along with other members of the founding crew will start off with a bit about why the conference came to be, and what their original vision was - and how it turned out.

Then we’ll hear from several other past organizers about how the event grew and evolved over the years.

Finally, we’ll wrap up with this panel-of-panels with the current organizing crew to share what went in to making this year’s event happen.

CSP (Cubesat Space Protocol) is the main open-source networking protocol for Cubesats: Rubik’s cube-looking miniature satellites gaining popularity due to their reduced cost and development time. This talk presents a vulnerability research deep dive into libcsp, the most popular implementation of the protocol. It uncovers a wide variety of vulnerabilities ranging from a Mitnick attack-like TCP connection forgery, insecure primitives and memory corruption. The talk will also cover the vulnerability disclosure process and its challenges.

Join me as I discuss what is likely the most famous examples of blinken lights. The CM-5/CM-500 super computer was featured in the background of Jurassic park due almost entirely to these lights. I was in the right place and the right time to take part in decommissioning one such machine.

In this talk I’ll briefly discuss thinking machines (the company), the connection machine (the super computer), and spend the rest of the time diving deep into the architecture of the board itself. Reverse engineering revealed surprising choices taken on a project where money is no object.

Ultimately, we’re left with the ability to display some graphics on the LEDs (I hope!) and having learned some interesting techniques.

In March 2026, people discovered Chipotle’s customer-support chatbot would happily solve LeetCode problems and write Python between burrito orders. Within days someone had turned it into a free coding agent running entirely on Chipotle’s compute. The internet laughed and moved on, and missed the actual lesson.

The uncomfortable part: nothing was hacked. No protocol cracked, no auth bypassed, no prompt injected. The “exploit” was a faithful reimplementation of the bot’s own chat client, wrapped in one thing, an OpenAI-compatible API which instantly dropped a captive support bot into the entire LLM tooling ecosystem: every SDK, every agent, every framework.

This talk names that pattern interface laundering, and argues it’s a latent property of nearly every AI feature shipping today. Your support bot, your “summarize” button, your semantic search: if it’s reachable and returns model output, someone can re-present it through a standard interface and take it somewhere you never intended. The friction you’re counting on as a control gets laundered away in about forty lines of code.

We’ll walk the (now-patched) Chipotle case, watch a deliberately vulnerable bot get laundered into a coding assistant on video, and leave with a hardening checklist that assumes the attacker isn’t a browser: where to enforce scope, why per-session rate limits are theater, and how to spot the “phantom client” whose absence of human fingerprints gives a proxy away.

Compiler optimizations can fundamentally change the structure and our understanding of how a binary runs under the hood, this makes binary analysis and reverse engineering much more difficult. One particularly important optimization is function inlining, where a compiler replaces a function call with the body of the function, usually stripping away the function name. This makes determining what function was originally contained in the program much more difficult.

This talk would explore our research into using Graph Neural Networks to detect function inlining in compiled binaries. We will examine how binary code can be represented in graphs, how compiler optimizations affects those representations, and how machine learning can be used to identify code that has been inlined.

The talk will focus on our approach and our results, while also providing the necessary background to understand our research. Our talk will explain why this is valuable for reverse engineering binaries, as well as identifying vulnerable functions and even finding CVEs. This is with the intent of furthering the cybersecurity landscape as currently detecting functions that have been inlined remains a problematic task. However, our method was able to detect inlined regions of a program 76% of the time. We will also touch on the limitations and future works of our research.

Attendees will leave with an understanding of why compiler optimizations complicate binary analysis, how GNNs can be applied to reverse engineering problems, and the applications that this work has to software vulnerability analysis, reverse engineering, and malware analysis.

Web Application Firewalls (WAFs) log millions of website requests every day, but turning that data into real protection is slow and manual. Security teams spend endless hours digging through giant log files to write blocking rules by hand—often days after an attack has already started.

In this talk, I will show a free, open-source Python tool I built to fix this problem. My tool takes raw security logs from Fastly’s Next-Gen WAF and uses machine learning to automatically group suspicious traffic by how it behaves, such as its timing or hidden patterns in the web requests. Once the tool spots a group of bad traffic, it automatically creates a ready-to-use blocking rule to stop the attack.

I will demonstrate live how the tool scans gigabytes of log data, catches a hidden botnet, and creates a working defense rule in less than two minutes. Attendees will also get free access to my code repository.

We all know Quora - or do we? Like Internet herpes, if you’re a Google user, it’ll follow you, forever, haunting your inbox with clickbait if you ever so much as click on a Quora result while logged into Google.

…But do we REALLY know Quora (or want to?)? Because, despite Quora being a pustule on the Internet that exists for the sole purpose of spamming SERPs, you’ll be shocked to learn that it’s also got - spoiler alert!!! - a dark, seedy underbelly. One that I uncovered while volunteering on an MP investigation. One that can yield surprising, disturbing and useful intelligence.

Sure, hating on Quora is low-hanging fruit (I’m short), but it’s got implications for OSINT and security overall.

In this talk, I’ll explain:

1. The toxic intersections of business strategy, platform policies, SEO, and socioeconomics that drive certain market segments to Quora in much higher numbers than you assumed
2. Why Quora benefits from tombstoning content (including content with minors’ PII) instead of actually deleting it
3. How to investigate Quora Quontent, which anyone can do, because it’s all served up completely legally in a grotesque tangle of GraphQL; hell, most of it is right there in the UI.
4. OWASP API and Mitre ATT&CK (R) CWE categorization for Quora’s approach.
5. What I think the OSINT community should do (hint: something. Anything!).
6. Quora’s silence on the issue. Well, I can’t actually explain it since it’s inexplicable to me, but I’ll totally describe that silence.

I don’t want this talk to end with pitchforks out and an angry mob marching towards headQUORters. I’d prefer a nice conversation of professionals generating ideas for remedying this Quorrible situation and a greater understanding of how to avoid disclosure of sensitive data when building with similar tools.

But pitchforks might be fun, too.

(Just kidding! I in no way condone violence or even the use of farming tools for anything other than old-timey farming!)

This talk was presented at DEF CON 34’s Recon Village (https://reconvillage.org/reconvillage-2026-defcon-34/talks/inquorigible-quora-in-missing-persons-osint). The original focus was on how to use Quora for missing persons investigations, but my research has since broadened; Quora can be used for MP investigations as well as investigating black hat marketing, psyops, “and much, much more!” from your nightmares.

Privacy policies are opaque documents, often loaded with obfuscating legalese, but they have real-world consequences for ourselves, our community, and the next generation. During this presentation we’ll explore the the current state of privacy as it applies to young people. We’ll discuss how schools factor in and how the data-broker industry is starting young. We’ll consider the real-world impacts of data breaches and how the people selling toys manipulate us when we’re just trying to let our kids play. We’ll wrap up with where we can go next, and how we can promote policy solutions to address this difficult problem.

“Think of the kids” doesn’t just have to be a political slogan, it can be a rallying call to make the world safer and more private for us all. Join the conversation to learn how to advocate for the young people in your family and our communities.

In the wake of all the AI news over the course of the year, you would think that the lowly Software Bill of Materials is not something you should pay attention to anymore. But as basic as the concept of an SBOM is, it is very difficult to implement in our modern software development lifecycle, and as such, difficult to realize the gains from it. This talk is about why you still need to pay attention to software metadata inventory and sharing, especially in the age of AI! We will see the security benefits of SBOM for a simple app, then watch as the complexity grows for our modern software stack. You will come away with some simple implementation tactics that will reduce the complexity of maintaining SBOMs while reaping the benefits.

Advances in quantum computing mean that Q-Day, when quantum computer can break typical RSA and ECDSA keys, may come sooner than previously predicted. Since those are the keys that protect essentially all HTTPS traffic on the web, maintaining privacy on the web will require an industry-wide push towards post-quantum cryptography. EFF’s Certbot and ISRG’s Let’s Encrypt are working hard to make sure everybody can participate in the post-quantum world as quickly, easily, and freely as possible.

This talk will discuss what that transition will look like specifically for those operating HTTPS web servers, and will cover the aspects of a TLS handshake that need upgrading: key agreement, certificate trust, and server authentication, with discussion of both public and private PKIs.

Security practitioners often need to quickly answer questions like: What telemetry records this activity? What fields should I investigate? What does the event actually look like? Has anyone already documented or detected this behavior?

This talk introduces a searchable catalog of security telemetry and more than 16,000 public detection rules. I’ll show practical ways threat hunters, incident responders, SOC analysts, SIEM engineers, and security engineers can use it to research unfamiliar events, identify useful telemetry, investigate suspicious behavior, and compare detection approaches.

Attendees will leave with several practical workflows they can immediately use during investigations, hunts, SIEM work, and daily security operations.

Imagine you live in a dystopian cyberpunk future in which the world is increasingly controlled by big tech monopolies working in lock-step with an increasingly fascist government. It’s not unrealistic to think that such a system might attempt to prevent its citizens from accessing certain written materials in order to prevent the spread of “unwanted” ideas. In the United States, “book bans” typically involve public and school libraries being required to stop stocking certain books. Books are not banned in the sense that they can still be produced, sold, and owned. However, there are countries that do ban and censor books today.

In this talk, I will show you how you can repurpose an inexpensive esp32-based WiFi-enabled light bulb and turn it into a cyberpunk dead drop used to wirelessly share books and materials your government doesn’t want you to see. Or you can host cat pictures. Or both!

OAuth 2.0 is the industry standard for delegated authorization. It allows a client to access a secured resource on behalf of a user without exposing the user’s password. Properly implemented, OAuth can provide robust security for your sensitive data. But done wrong, your application could be vulnerable!

One such vulnerability is an interception attack. This is where a malicious actor can intercept an OAuth token over the network. The token can then be used to impersonate a client, giving the actor access to a resource they’re not supposed to have.

In this talk I’ll first give a high level overview of OAuth. Then I’ll demo how to perform an interception attack on an application which has implemented OAuth insecurely. Then finally, I’ll show how to secure an OAuth authentication flow to make an interception attack difficult to pull off.

Attendees will leave this talk with a high level understanding of what OAuth is, where it can be vulnerable, and how to implement it securely in their applications.

Someone at your company just prompted an app into existence. Someone outside your company can do the same with your logo and a login form. Neither needs to register a domain. So how are you supposed to find either of them?

I started with the platforms. VibeCheck grew into more than 176.5 million observed hostnames and 100,000+ screenshot captures across AI builders and the publishing infrastructure around them. We’ll get into Base44, Lovable, Bolt, Emergent and friends: how we found their app hostnames, what DNS fingerprints they leave behind, and how to follow those clues onto custom domains. I’ll also show how we put a bunch of laptops in my home lab to work, splitting discovery, OCR and screenshot review across an agent swarm.

Once you start looking at the pages, you find things a company probably wants to know about. Clinical caseboards linking names to procedures. Loan-approval screens displaying borrower details and credit decisions. Administrator usernames and passwords printed beside the login form. Is this real business data? A demo? Something somebody meant to share? We’ll work through what helps us tell the difference, where those clues get shaky, and what we’d need to check next.

Then there’s the abuse: corporate and mailbox login clones, billing forms asking for card details and Social Security numbers, verification pages requesting phone codes, wallet download and recovery lures, and pay-first offers promising a bigger payout. A Ledger support form asks for recovery words while warning never to share them. An FBI-branded “support center” claims verified government status and sends visitors to Telegram. We’ll examine what these pages ask people to hand over—and compare them with legitimate brand-controlled assets using the same platforms.

I’ll show how we found these, why they deserve attention, and what we’d check next. You’ll leave with a workflow you can use yourself, free OSINT starting points, and historical and reverse-DNS pivots that help move an investigation forward. I’ll also share a large screenshot library of some of the worst examples, with capture dates and investigation notes, so you can dig well beyond what fits into 20 minutes.

We tracked an Android malware campaign across multiple applications published on the official Android marketplace, including SMS messengers, games, PDF scanner, and phone-cleaning utilities. Although the apps appeared unrelated, each contained the same randomly renamed loader responsible for decrypting and executing a hidden payload.

Over time, the operator constantly changed implementation details: payload execution methods, C2 protection mechanisms, packaging strategies, developer accounts, and application themes. In several cases, the dropper was embedded as a seemingly legitimate third-party SDK. These changes created the appearance of new malware families while the underlying operation remained the same.

This talk walks through the investigation process used to connect the samples and uncover the campaign. We will show why analyzing only the newest sample can lead to fragile detections when malware builders rotate techniques instead of evolving linearly, how common static-analysis workflows missed critical indicators, and which pivots ultimately revealed the relationships between samples.

Attendees will leave with practical hunting and attribution techniques that can be applied to Android malware investigations, including package structure analysis, fuzzy-hash correlation, signing-time clustering, and methods for distinguishing meaningful pivots from misleading ones.

Everyone assumes AI speeds up security review by automating the scan and the checklist. That’s not where the time was going. The real bottleneck is a scarce human reviewer and the back-and-forth of iterating a design until it’s actually reviewable. This talk is about what happens when you stop trying to make the machine part faster and instead spend the AI’s time absorbing the iteration rounds that used to eat the human’s with real numbers from an internal review pipeline.

Just 12 months ago, 5% of pull requests at my company were authored by an agent. Today, that number has soared to over 95%. Maybe that’s staggeringly high? Or maybe some are wondering why that number isn’t 100%. One thing is for sure: as tools become better and token budgets balloon, engineers (and even non-engineers) are using agents to ship exponentially more code than ever - for better or worse.

When humans become more disconnected from code with their name on it, how do we create incentive structures that facilitate safe and sustainable development? How do we protect ourselves from endless and seemingly plausible slop? How do we secure an environment where code volume outpaces human attention? This talk shares successes, failures, and field notes from the front lines of security engineering in this new age of software.

Valve Software is an award winning video game developer, publisher, and digital distribution company known for both the quality of their first party titles and the widely popular distribution platform Steam. With a commanding share of the PC digital distribution market, its presence across Windows, Linux, and OSX machines is ubiquitous.

In this talk, we’ll dig into the technical details of the platform, from its distribution and authentication mechanisms to its management of multi-user environments and privileged digital assets. We’ll briefly detail the architecture necessary to support Steam Remote Play and Big Picture modes, their remote attack surface, and the quirks embedded within each. Finally, we’ll be detailing several local and remote vulnerabilities discovered and exploited, and exploring the design choices that ultimately led to broken trust boundaries.

Frontier models are powerful, but questions around data privacy and terms of use are real concerns in the security community. I started as an AI skeptic and ended up building and self-hosting my own agentic framework on a $280 used GPU that continues to power my personal cyber security research. This talk covers the projects and lessons learned that got me there. In addition to the amusing anecdotes, I’ll go over gotchas from self hosting, strategies I implemented to overcome challenges with my exclusion of frontier models, and the many times frontier models were confidently wrong. I did not reject frontier models outright, I am using them to build the infrastructure I need to continue my research without worrying about data exposure or terms of use violations.

A Minneapolis data broker lost my name, SSN, and date of birth in the MOVEit/Cl0p breach, notified me 42 days late in violation of Washington’s 30-day law, then refused my $1,050 invoice for the remediation time their negligence cost me. I filed in small claims, beat their corporate jurisdiction motion, and the judge ruled that a security professional’s time assessing a breach is compensable at professional rates. They paid. This talk is the replicable playbook: a $50 filing fee, a few hours of prep, no lawyer. Class actions pay lawyers. Small claims pays you.

Border Gateway Protocol (BGP) is the foundation of Internet routing, enabling networks around the world to exchange information about how traffic should reach its destination. Yet BGP was designed around trust, and that trust can be exploited.

In this talk, I will introduce the fundamentals of BGP, explain how route advertisements influence Internet traffic, and examine what happens when those advertisements are incorrect or malicious. We will explore several notable cases of nation-state-sponsored BGP hijacking, including incidents affecting services such as YouTube, Apple, Telegram, and others.

The talk will then examine the mechanisms used to detect and defend against BGP hijacking, including the role of routing security practices and the challenges involved in deploying them at Internet scale.

Finally, I will present a short case study examining university network prefixes and their exposure to BGP hijacking. Using an experimental methodology, we analyze the likelihood of several university prefixes falling victim to a BGP hijack. Specifically we’ll examine: University of Oregon, Portland State University, Utah State University, Stanford University, and others.

Cloud VMs routinely bootstrap powerful identities through the Instance Metadata Service (IMDS), where network reachability can become implicit evidence that a caller is entitled to the hosting VM’s identity. Real-world metadata incidents show manifestations of this trust assumption through SSRF and privilege escalation risks. This talk focuses on a less obvious case: hostile sandboxed workloads. As AI agents and other untrusted workloads are increasingly executed inside cloud sandboxes, ensuring that these execution boundaries also enforce identity boundaries becomes critical.

The talk demonstrates how a Hyper-V isolated container running a hostile workload can acquire the hosting VM’s managed identity through the host’s networking path to IMDS—even when that workload is protected by a nested virtualization boundary with an isolated kernel and additional in-guest security controls are present. The talk examines why common mitigations such as networking controls, firewall rules and HTTP header requirements, while providing defense in depth against traditional metadata attacks, lose effectiveness for sandboxed nested workloads. Furthermore, the talk demonstrates the sandboxed workload scenario to showcase how traffic attribution is challenging and why simply blocking access to the metadata endpoints per-sandboxed container does not scale.

Finally, Azure Metadata Security Protocol (MSP) is presented as an example of a protocol-level redesign that replaces implicit trust in network reachability with explicit authentication and authorization. MSP combines transparent eBPF interception, secure key establishment, deny-by-default authorization, and HMAC-based request authentication to IMDS. The design demonstrates how Zero Trust principles can be applied to identity bootstrapping while preserving transparent access for existing applications and establishing a stronger identity boundary around sandboxed workloads in the cloud.

SaaS products have long since become necessities for the modern enterprise, even the most restricted environments commonly include exceptions for outbound network connectivity to these essential services. But when any scoundrel off the street can sign-up for a cloud tenancy is it safe to trust the platform as a whole? Explore the risk hidden in multi-tenant SaaS platforms by example: a command and control channel built on the Datadog platform. By boiling down application features to their basic building blocks we can find creative ways to fulfill the minimum contract necessary for C2 communications.

Bring Your Own Vulnerable Driver (BYOVD) is already well understood: attackers load a legitimate signed driver and abuse its privileged capabilities. This talk presents Bring Your Own Trusted Caller (BYOTC), a variant that can be more far-reaching than a single vulnerable-driver bug. Many drivers restrict powerful operations to a trusted user-mode client; if an attacker can control or impersonate that client, the driver’s authorization check becomes a confused deputy. I have confirmed this pattern across several products, and the same design could affect dozens more drivers.

I have published the first installment of my BYOTC research (https://xusheng.dev/posts/byotc/main/), which examines Malwarebytes Chameleon and System Informer. In this talk, I will present additional cases: a related process-hardening issue in VirtualBox; dynamic attestation replay in Avast; creator-controlled trusted clients in 360 Total Security, Huorong, and Antiy ATool; and a basename-based caller gate in HitmanPro.Alert. The cases show different ways an attacker can cross the same trust boundary, with demonstrations scoped to the operations each test actually proved.

I will close by examining Microsoft’s preview Windows Endpoint Security Platform (WESP). Moving antivirus enforcement out of the kernel can reduce some driver risk, but it does not automatically authenticate a client to a privileged service or preserve that client’s runtime integrity. The BYOTC question remains wherever a privileged component trusts a client that an attacker can influence. Attendees will leave with a method for reviewing that boundary in both drivers and user-mode security services.

Ransomware groups have historically relied on well-understood encryption to lock victims out of their data. But a new generation of families is deploying cryptographic schemes of surprising sophistication, far exceeding what’s needed to prevent decryption. The Kyber ransomware combines hybrid post-quantum key encapsulation mechanism alongside elliptic curve key exchange for symmetric key derivation, seeded by multiple independent entropy sources. It is designed to resist quantum computers that don’t yet exist. CRY0 implements an uncommon authenticated cipher paired with a dual-layer asymmetric key wrapping scheme using a modified Chinese national cryptographic standard, that enforces mutual distrust between operators and affiliates, preventing either party from decrypting files alone. A third family is AiLock, which explicitly markets itself as “AI-assisted” and wraps file keys with a different post-quantum algorithm, yet ships with implementation flaws that suggest the developers may not fully grasp the cryptography they adopted. This level of implementation complexity, and the mistakes they incur, would have been impractical for most malware developers just two years ago, which makes us question whether AI-assisted development tools are enabling threat actors to reach beyond their traditional capabilities.

This talk dissects the encryption architectures of these families, analyzes why this complexity is operationally unnecessary, and investigates the indicators of the AI-assisted development hypothesis. It then turns to the defensive side: the author shares a practical field report on using LLMs as reverse engineering co-pilots against these same families, documenting where they proved invaluable and where they consistently fell short. The emerging picture suggests an asymmetry worth examining: current AI tools may lower the bar for building complex malware more than they lower it for analyzing it.

Security analysts do not have a shortage of information. Vulnerability findings, threat intel, compliance requirements, audit evidence, and a growing stack of SaaS tools all compete for the same finite attention. Somewhere in that flood we’re still expected to make sound risk decisions.

This talk skips the “should security use AI?” debate and ask a narrower, more useful question: can AI make the analyst faster without becoming the analyst?

Drawing on hands-on security work, I will walk through three real use cases where AI has meaningfully sped up my process bit have deliberately kept it out of the driver’s seat:

Vulnerability exposure research - using AI to structure the investigation into exploitability and blast radius, not to declare severity

Compliance analysis - using AI to develop compliance requirements into evidence, questions, and gaps to cut audit evidence gathering time.

Shadow AI / SaaS risk review - turning a pile of vendor documentation into a structured set of data-handling and third-party processing questions

Each example follows the same workflow: research, correlate, identify what is missing, analyze, verify, decide, and document. AI can take on the analyze and verify work while not touching the verify or deciding.

This talk also covers where this AI is less useful. For example, AI’s tendency to sound confident while missing context or why a polished wrong answer can be more dangerous than an obviously wrong one. Attendees will leave with a practical framework for where AI genuinely helps a security analyst’s workflow and how to keep human judgment in the loop without losing the speed gains.

This is not an AI-evangelism talk or an AI-skeptic talk, but a practitioner’s field report on where the line actually sits.

Notes for reviewers: All examples come from real practitioner work; vendor names and specific findings will be genericized or represented rather than shown as confidential material. Demos use recorded/screenshotted walkthroughs where live data isn’t

Ever looked out over the raft of ocean-boiling datacenters and ever-rising token costs and thought to yourself “I mean, I’ve got a home lab and a wallet to set on fire… how hard can it be?” Well, I sure did.

Tune in to learn what I did with self-hosted and open-sourced AI tool-chains. You won’t leave an expert but you will know where to start and what not to do.

Infosec’s community-maintained lists have become a backbone of security. Lists come in many forms, but share a common ethos and grounding simplicity:

1. Cyber intel is a team sport
2. Do The Basics well

Regardless of the specific campaign, actor, or hype cycle, community lists are still an effective and deterministic full-stop in kill chains.

In this talk we distill real-world IOCs to their core service, relate to public lists, apply them to kill-chains, and apply low-regret methodology to gain leadership buy-in.

Attendees will be provided the lists, reference material, and techniques for a successful path to deployment of living off the land defenses.

Small and Medium-sized businesses (SMB’s) are under the same threats and risks as large companies, yet they have tiny budgets and limited staff to support infosec and compliance programs.

The frequency and sophistication of attacks is still increasing. SMB’s make up the critically important economic and services Mainstreet of America. They need better solutions than the current typical infosec program and tabletop exercises used today, which were designed for the military and Fortune-500 companies.

What if we could redesign infosec development processes for SMB’s that would lower costs and double or triple the yield of value and results? What if we could achieve those results in less time and at less cost than a typical program?

Impossible? Not if we completely change the organizing principle that drives the program. It’s time to level-up the human skills and training side of infosec.

Come see how it is done.

When an AI-assisted query nearly caused special forces to board a commercial vessel based on false reports of nuclear cargo, commentators blamed model hallucination. In reality, this was a structural pipeline failure: the system borrowed the format and authority of an official brief without carrying provenance, confidence, or observation timestamps through the orchestration layer. This talk presents a practical architectural pattern (featuring a three-field claim schema, strict display rules, and a human feedback loop) that ensures machine-generated findings carry verifiable receipts before humans act on them.

We already know that LLMs can catch bugs, but what about fixing them autonomously, at scale, without disrupting production?

To investigate this hypothesis, this talk puts together an illustrative but very common scenario: a system that relies on a vulnerable library to handle untrusted user inputs across multiple service instances. By using canary deployments with built-in sanitizers, the system self-monitors for possible crashes and then, once a crash is identified, it feeds the dump into an LLM analyst harness. The harness then proceeds with root-causing the problem, implementing a patch for it and finally wrapping the patch into a live-patching unit. Libpulp, a userspace live-patching framework, completes the pipeline by applying these fixes across running production instances without service downtime.

While the proposed pipeline system exists as a case-study PoC developed specifically for the talk, its development leveraged established open-source components such as libpulp, debugging tools and compiler-based sanitizers. In this talk, we plan to dive into each of these specific components, discussing all the technicalities around how they work, why they matter, and how all the steps can be connected and automated through the usage of LLMs, achieving a self-healing system.

I buy boxes of old floppy disks (3.5 and 5.25) off eBay and out of thrift stores, media people sold, donated, or threw away decades ago, sure that whatever was on them was long gone. It wasn’t. Across 7,000+ disks I’ve recovered nearly a quarter-million files. 1/4 way through the 20,000 in the garage.

This talk walks through the workflow I built to capture, extract, and analyze that archive at scale — and what it surfaced: personal records and sensitive PII, decades-old viruses still sitting on the disks, and a startling amount of private adult content. You’ll learn how cheap, open-source hardware and an automated, ML-assisted pipeline turn forgotten media into searchable data — and why you should destroy your old disks before you ever give them away.

Four years ago, Delete Your Data came to BSidesPDX to talk about reducing your online footprint and exercising your privacy rights. Since then, AI has changed the threat model, and we’ve rebuilt our approach around agents that help protect you from other people’s agents.

Our founding principle is that everyone should have a meaningful say in what happens to their information. Or, as the hipsters say, “Consent is sexy.”

Delete Your Data helps you ask companies to remove your information on your terms. A lawyer reviews the request, the company’s response, and the available evidence. If your information shows up again after a company says it has been removed, our sharky litigator friends will help you decide what to do next.

We’ll show how agents can help find exposed information, prepare and track deletion requests, follow up with companies, and organize the evidence. We’ll also discuss how dashboards and simple interfaces can help people apply human judgment quickly. After all, an agent can assist you, but it cannot make your decisions for you.

Attendees will leave with a practical framework for evaluating deletion tools, deciding what to automate, and identifying where human review belongs. We encourage everyone to be proactive about their data hygiene, whether or not they use our tools.

Within physical social engineering, exploiting assumed trust is key. You need to dress the part, know the lingo, and make anyone who looks your way think you’re simply an employee from out of town or a trusted vendor. Over the past couple years, Lucas Zahorik and Tyler Aldous have compromised the physical security of banks, law firms, hospitals, and fortune 500 companies, all in the name of testing and improving security. This presentation will cover assumed trust and gaining entry to a site, illustrated with a war story of how they compromised a company utilizing fortune cookies left in their break room by combining physical entry techniques and advanced phishing.

Open-weight models have grown quickly in popularity in 2026. The shortage of DGX Sparks and the huge number of orders for the new Mac Studio M5 Ultra show how much people want control over the models they run. But running a model yourself means trusting weights you did not train.

In this talk, we will introduce a malicious behavior into an LLM: when used in a certain context, the model exfiltrates sensitive data. We will then explore techniques to identify whether a model has been tampered with to add such a backdoor, and how to analyze it.

Earlier this year, a contributor tried to smuggle North Korean malware into Better-Auth, the popular JS auth library, through pull requests that shipped real, requested features. Buried between the legitimate changes sat an obfuscated payload in the build config. It fires the moment anyone runs npm run dev, pulls encrypted second and third stage payloads from blockchain transactions, and spawns a hidden process that keeps running after your build exits with code 0. We caught it, reverse engineered all three stages, and found the same malware signature in 30+ other repos on GitHub.

This talk is a live teardown of the full attack. You’ll see how the malicious code hides off-screen in GitHub’s default diff view (try scrolling right sometime), how a 37-character scrambled string decodes to constructor so the payload can build executable code that no static scanner catches, why hosting malware on a blockchain means it can never be taken down, and what the C2 server steals first. Spoiler: OPENAI_API_KEY is near the top of the list, right next to your AWS creds and GitHub token.

LiteLLM and axios were hit by supply chain attacks the same week. If you ship on open source, this pattern is already aimed at you. You’ll leave knowing exactly what to grep for in your own dependencies.

Modern (detachable) camera lenses are engineering and optics marvels with light speed autofocus and incredible optical qualities. To enable this versatility, most camera systems no longer have lenses with manual controls. Where a vintage lens can be repurposed or adapted to various camera systems with a simple adapter, modern electronic lenses are useless when detached from their particular camera.

Long ago, the author has settled onto Sony camera ecosystem and has amassed a collection of excellent lenses but their proprietary camera-lens communication protocol limits their use. By employing common hardware and software reverse engineering tools and techniques, I was able to reconstruct the signaling, timing and messaging details of this protocol and implement a controller that behaves like a camera does. This makes it possible to use these fancy lenses in custom DIY project such as a Raspberry Pi based camera.

I will present the tools and techniques employed, stumbling blocks encountered and results achieved. I will be open-sourcing an implementation of the protocol that enables focus and zoom control of Sony e mount lenses via python and C via GPIO.

Scammers prioritize speed and human interaction; our solution prioritizes automated distraction. This talk details a novel adversarial defense strategy: deploying Large Language Models (LLMs) to functionally waste the time of fraudsters. We begin by analyzing the history of scam tactics and the mechanics of scambaiting, establishing the threat landscape. We then dive deep into the implementation: a practical proof-of-concept system built with Ollama and Go that automates the engagement and confusion of fraudulent communications. Crucially, we demonstrate that powerful defense can be resource-efficient, showing how this agent was successfully deployed on a modest stack, an email service and a laptop with a modest GPU. The presentation culminates in the results, showcasing the hilarious outcomes that prove our AI successfully turns the tables, ensuring scammers waste their time on digital nonsense instead of real victims.

Most attempts to automate code auditing with LLMs follow a familiar pattern: provide more context, add external tools, and ask the model to produce findings. While this approach can generate vulnerability reports, it often struggles to produce repeatable and trustworthy security assessments. This talk presents lessons learned from building an LLM-assisted security auditing pipeline and the architectural shifts that emerged when common assumptions failed. Larger context windows reduced reliability, verification collapsed when discovery and validation were combined, and agent-style workflows frequently produced plausible but difficult-to-verify narratives. The resulting primitive-first approach decomposes analysis into structured evidence objects and recomposes findings from independently validated relationships, reducing inference, improving traceability, and enabling more trustworthy AI-assisted security reviews.

AI assistants summarize their work at the end of handling a prompt, but they often omit important details. Even worse, they can be tricked into hiding details when they’ve been hacked. I wanted a tool that deeply monitors process, filesystem, and networking usage when my assistants work. That’s why I created Dex, a POC auditing tool, which makes it easy to understand what my assistants actually did between prompts. Dex is like Little Snitch for AI, but it goes further, including deep L7 network inspection of model and tool requests, child process tracking, and filesystem access. In this talk, we’ll use Dex to take a closer look at how AI agents work under the covers, including process and networking mechanisms, to see how agents can disguise their work while leaving an audit trail if you have the right monitoring tools in place.

Some of the most prominent conversations I hear from new grads and current students is, what do I focus on, how do I get a job in offensive cyber security, and why is everything on fire. We don’t have time for that third bullet point, but I plan to talk about the first two through the lens of what I would do if I had to start fresh today.

Developers constantly live in fear of CI/CD supply chain attacks or vulnerabilities causing damage in an unrecoverable way. The real question isn’t, “was my repo affected?” instead it’s, “what could the attacker actually gain, or reach given the affected workflow run’s token scope and permissions?” Currently, a developer does this manually by checking permissions across all repositories case by case and analyzing the impact. This talk presents a way to automate the process by introducing a tool that checks the exact token scope and permissions that run had during the compromise window, checks what’s reachable, identifies exposed secrets, and determines whether lateral movement to other workflows or repos is possible or not. After analyzing it will list the impact with risk score and remediation to prevent the blast radius from further increasing.

As long as people have been talking on two-way radios, other people have wanted to listen to them. But where do you begin? Radio comms rely on practices and jargon to optimize communication in a domain where you can’t transmit only to the person you want to talk to, you don’t know if your transmission can reach them, or if their transmissions can reach you.

If you simply eavesdrop on people talking on the radio, you’ll probably have no idea what they’re talking about. This workshop will first teach you how to effectively communicate on the air so that you can understand the transmissions that you will also learn to intercept — within the bounds of the law.

You will learn the basics of handheld radio operation, radio language and etiquette, and use them to actually communicate on the air. Then we’ll cover how to identify frequencies used and intercept them using low-cost, off-the-shelf equipment, including directional antennas and software-defined radio (SDR). Finally, we’ll explore the limits of radio interception, including dealing with trunking systems and encrypted comms.

This is going to be a cool as hell workshop where we will learn by doing. No radio experience is expected or required! Folks who have never so much as held a walkie-talkie are strongly encouraged to participate. You just have to be ready to go outside — rain or shine!

The world has changed…

It began with the forging of the modern enterprise. To the cloud were given vast kingdoms of compute. To the mobile apps, a presence in every pocket. To the machines that now think, knowledge stolen from the creators. And to bind them, to let each speak to each across the dark spaces between systems, the architects forged the APIs.

But in the forging a flaw was hidden. For an API answers any who ask it correctly, and it does not always stop to learn who is asking. So it came to pass that buried among ten thousand harmless calls there waited a single request. One request with dominion over all the others. One request to read what was meant to stay hidden, to act in the name of any user, to claim the powers kept for admins alone.

One Request to bring the whole realm to ruin, and in a single call, destroy it.

Game-based learning can be fun. You may (or may not) have heard about Cyber Defence Dice or Security Cards. In this interactive workshop, attendees will get hands-on experience with our new security cards - an extension that bridges threat discovery and defense. Teams will allocate scarce budget and re-evaluate residual risk on a matrix.

https://blackhoodie.re/BsidesPDX2026/

When you think of robot vision you might think: collect a dataset, label thousands of frames, train a model, and run it on a GPU. Nope! The gate detector that qualified my drone for an autonomous racing competition was plain OpenCV on a single camera feed, and I am going to teach you how to make it in this workshop!

If you’ve never written a line of computer vision code, you’re the person this course was built for!

Together we will cover: Segmentation, Lighting Invariance, Target Selection, Partial Gates, and Confidence Ranking. Each stage presents a new problem for us to solve and eventually our code will be able to handle real drone footage from an actual flight.

Everything you need is provided in a Docker container. No GPU needed. A laptop from the last decade will work. You’ll leave with a detector you wrote yourself, a graded test set to keep improving it, and the first steps toward flying a drone on vision alone!

CTFs. The thrill of getting a flag is unmatched, but the narratives tend to be … flat. But they don’t have to be! Play an analog-horror, ARG-style capture-the-flag set in the heart of Portland. OSINT, cryptography, web application vulnerabilities unlock a story about how a fictional man went missing (mostly) without a trace.

A bit of background regarding the story and infrastructure will be shared before opening it up for live play. The difficulty ranges from easy to medium but experience is not necessary (non-cyber participants were able to solve this as well)! Participants learn practical OSINT, cryptography, network enumeration, and web-security techniques by solving an interconnected narrative CTF.

BSidesPDX is all about nurturing the local infosec and hacking community, and part of that is supporting folks as they step onto larger stages. Have you attended DEFCON, or have you presented at BSidesPDX, and are you wondering what that next step looks like? We’re all recent first-time speakers at DEFCON. We’ll tell you a little bit about what we presented and what the process was like, but then open the floor to questions from you. The session will be off the record and not recorded to make sure we get the honest questions and honest answers.

BSidesPDX is all about nurturing the local infosec and hacking community, PDX Hackerspace has lots of overlap and a similar mission. Have you wondered what it takes to build and support the community, and keep the lights on at the hackerspace?

We’re all members of and volunteers at CTRL-H/PDX Hackerspace. We’ll tell you a little bit about our interests and why we’re involved in the space, but then open the floor to questions from you. The session will be off the record and not recorded to make sure we get the honest questions and honest answers.

BSidesPDX is all about nurturing the local infosec and hacking community, and part of that is supporting folks as they advance professionally. Are you a student looking ahead toward graduation and a job in the field? Are you looking to change fields into infosec? Are you a manager that has to hire entry level positions? We’re all recent additions to the infosec workforce. We’ll tell you a little bit about what we do now, but will focus more on what we wish we knew before we started. We’ll open the floor to questions from you. The session will be off the record and not recorded to make sure we get the honest questions and honest answers.

Brought to you by PSU and OSU alum, it’s another year of CTF! Bring a laptop and some friends, there will be challenges for all skill levels. The competition will run all conference long, so you can play in between talks. We hope to see you there!

Categories: web, rev/pwn, crypto, and social engineering Prizes: TBD