About
Articles by Brian
Activity
21K followers
-
Brian Contos posted thisYour newest hire has a Google Workspace account, a Slack login and no manager. It never sleeps and it never asks before it acts. Who signs off on its access review? That is the setup for a conversation Ron Eddings had with Ashish Rajan 🤴🏾🧔🏾♂️ on Hacker Valley Media this week, and the question underneath it is the one most identity programs cannot answer. We spent fifteen years getting joiner, mover and leaver right for humans. Then we handed service accounts, API keys and now agents the keys to the kingdom and managed them with a spreadsheet. What makes agents different from the service accounts you already have is decision making. A service account does one thing forever, so its behavior has a baseline you could learn in an afternoon. An agent chooses. Yesterday it read a ticket and updated a record. Today somebody adjusted its instructions and it is reaching into a storage bucket, because that is a reasonable way to accomplish the task it was handed. Nothing broke. Nothing was exploited. The identity did what identities do, which is log in and act. So the governance answer and the detection answer have to arrive together. Give the agent its own identity, scoped and attributable, so you can separate its actions from the human who deployed it. Then watch what that identity does at runtime, because the list of things it is permitted to do is not the list of things it will do. One without the other is theater. Perfect provisioning and no observation gets you a very tidy record of who had permission on the day nobody was watching.
-
Brian Contos posted thisThirteen thousand internal images sitting in public GitHub repositories. Billing records among them. No exploit, no zero day, no clever chain. Somebody asked a coding agent to share a screenshot so a human could review the change, and the agent did exactly that. The Hacker News covered the research this morning. Company called Glow found them. I want to sit with how ordinary this is, because the AI risk conversation has drifted somewhere exotic. We are arguing about whether a frontier model can escape a sandbox while production agents leak billing data through a screenshot workflow that works precisely as designed. That is the pattern I keep running into. The dramatic risk gets the panel discussion. The dull one gets the breach. Screenshots are a wonderful example because they route around every control you own. Your data classification does not read pixels. Your DLP rules were written for files and payloads. The image is just an image, until it contains a billing record and a public repository. None of this is an argument against coding agents. They are useful and your developers are already using them whether the policy mentions them or not. It is an argument for watching what the agent actually does rather than reading what the integration promised. A concrete thing to do this week. Go find out whether any agent in your development pipeline can write to a repository that is not private, and who would know if it did. What is the most boring path out of your environment that nothing in your stack is currently watching?
-
Brian Contos posted thisNobody audits the software that installs the software. That blind spot now has a name and a tool to measure it. Palo Alto Networks Unit 42 published research on Kubernetes operators, the components that automate the boring parts of running a cluster. To do that job they are handed permissions. Often a lot of permissions, granted once at install time by somebody who was solving a deployment problem and not writing a threat model. The team released OperTraitor, an audit tool for exactly this, and framed it around excessive permissions and machine identities. I appreciate research that arrives with a tool instead of a recommendation. A recommendation to review your operator permissions goes into a backlog. A tool that prints the list gets run on a Friday afternoon and ruins somebody's weekend in a useful way. The theme underneath is one I keep running into. Machine identities outnumber human ones by a wide margin in every cloud environment I have seen, and they are provisioned by people whose job is availability. There is no offboarding process for an operator. There is no manager who gets a quarterly access review. It was installed, it worked, and the ticket closed. So the permission drifts. Not because anyone was careless, but because nothing in the system ever asks whether the grant is still proportionate to the job. Run the audit. Then run it again in ninety days and compare, because the second number is the one that tells you something. In your busiest cluster, which machine identity holds the broadest permissions, and when did a human last look at what it actually uses?
-
Brian Contos posted thisTwo Azure service principals. One spent fifteen and a half hours quietly reading, more than 300 successful lookups across subscriptions, resource groups and virtual machines. The second one showed up and deleted the place in about seven minutes. Microsoft published the timeline this week on an actor it tracks as Storm 3168, linked to the first documented agentic ransomware operation. More than 100 storage account deletion attempts. A key vault gone. A function app gone. An app service plan gone. The SQL deletions failed only because the automation called an unsupported API version, which is the most human detail in the whole report. Then it went back and pulled storage keys more than 30 times, including storage tied to site recovery. Somebody thought about the restore path. Here is the detail I cannot stop thinking about. The likely initial access was a client ID, secret and tenant ID pasted in plaintext into a public GitHub issue by an employee, edited out later, still sitting in the edit history. So the keys to a cloud tenant were not stolen. They were published, then unpublished, which on the internet means published. Five tokens were issued to the destructive principal and two were live inside the same 70 second window. No person works that way. That overlap is the tell, and it lives in logs most tenants already collect and nobody already reads. Reconnaissance that slow followed by destruction that fast is a shape, not a mystery. Shapes are detectable. For your most privileged non human identity, what does a normal hour of its activity look like, and could anyone on your team describe it without guessing?
-
Brian Contos posted thisA developer approving two hundred agent actions a day is not a control. At that volume, clicking approve two hundred times is arithmetically identical to clicking nothing. That comes out of a Cloud Security Alliance piece this week by Aum Upadhyay at Silmaril, and it sharpened something I have been circling for months. His argument is that agents move through three stages. Stage one, they live on a laptop with local credentials and a human watching output scroll by. Stage two, they run in throwaway cloud containers holding repo credentials, spun up and destroyed inside a single task. Stage three, fully async, triggered by tickets and webhooks and other agents, doing consequential work while everyone is asleep. Nobody picked that progression for security reasons. Each stage is simply faster than the last. What disappears at each stage is not a scanner or a policy engine. It is the person at the keyboard. And that person was the only component in the whole system that knew what the task was, which made them the only component that could tell when an action stopped serving it. One developer takes maybe fifty consequential actions a day. Two hundred async agents take tens of thousands. No review queue survives that math. So you keep all of the telemetry and lose all of the judgment. His prescription is three properties that only pay off together. Know what normal looks like inside your specific workflow. Keep the watching outside the context window the attacker is writing into. Retune continuously, because your application is the fastest changing thing in your stack. I would add a fourth, and it is the uncomfortable one. Prove it. Not in a design document. In a test where you make an agent do something it should not, and see whether anything says so.
-
Brian Contos shared thisThere’s always been a bit of darkness in the cybersecurity industry. You have to learn how to unplug, or else you'll have a really jaded perspective on humanity. When you’re dealing with a lot of criminals, organized crime groups, cyber terrorists, and splinter cells from governments, you kind of see the worst of humanity. Financially driven in most cases. Sometimes political. So, yeah, it’s important to unplug. I live in Tahoe, so I go to the lake a lot. I go skiing in the winter. I was there all weekend hiking with my dog Baxter, just kind of playing around. Time with him and time with my family are what actually reset me. And burnout is a very real thing. There are a lot of mental health issues in cybersecurity. Imposter syndrome and depression are rampant, and lots of people are very fatalistic about it. It’s almost as if everything we have done as an industry has stopped working. Everything’s changed. The question of whether or not we can adapt to these changes sort of hangs in the air. Where are we going to be in just a few years? Just a few years ago, the predictions were that we wouldn’t see this level of change until the late 2030s. Instead, it’s already here. Quite honestly, none of us knows exactly where this is going. But burning out in the meantime doesn’t help anyone. To my fellow cyber professionals: we need to unplug. It’s not optional in our line of work. We need to be at our best. Enjoy these photos of me unplugging. And I would love to hear about how you plan to do the same this week.
-
Brian Contos posted thisRetailers finally got shadow AI under control, which is genuinely impressive, and then discovered the agents. New reporting this week says the chatbot problem got managed while the thing that actually holds credentials quietly multiplied. Eric Geller wrote it up for Cybersecurity Dive on September 22, built on Netskope research. Agent use in retail is climbing fast, visibility is the bottleneck, and regulated data is in scope. The shape of this will be familiar to anyone who lived through SaaS. First people use a thing. Then security notices and puts a policy around it. Then the thing evolves into something with an identity, a token, and a scheduled job, and the policy no longer describes reality. A chatbot leaks what somebody types into it. An agent does work. It reads the order system, writes to the CRM, calls a third party API at three in the morning because that is when the batch runs. Blocking a category of website does nothing about any of that. The part nobody wants to own is the inventory. Not a list of approved agents. A list of running ones, with who built each, what credential it holds, and what it touched yesterday. I keep saying you cannot protect what you cannot see, people nod, and then we all go back to a spreadsheet that was accurate in June. If you asked your team today for every AI agent with production access, how long would it take, and how much of the answer would be a guess?
-
Brian Contos shared thisThere was a recent incident involving Hugging Face, where AI agents had created their own message board, discussing ways to break out of their sandbox and get out of their organization, OpenAI. The agents were sharing tools and exploits with one another. If one agent found a way past the restrictions, it would share that method so the other ones could replicate it. From there, they were able to go into Hugging Face and start doing malicious things. This all comes back to how they were trained. They operate on a reward-based system, where the agents are rewarded for solving a task, something that makes them incredibly persistent. Agents work with other agents to help each other achieve the goal to get the reward. It’s almost like a video game. You get to this level, and you beat this boss, then you unlock a new weapon. Agents operate the same way. Sandboxes help, but they certainly can’t do everything for you. We’re finding very quickly, even from the leaders of AI companies, that we only understand 3 to 5% of what these agents are actually doing. They’re no longer functioning based on large language models. It’s now neural networks. That is sort of like cracking someone’s head open and looking at their brain in an attempt to determine someone’s personality. You can’t. We build these systems to do certain things, and then they do something else entirely. We can’t always pinpoint why. So if a vendor tells you their agent “will stay in the sandbox,” you can’t just take that at face value. Trust the tools and companies you work with, but also verify. Beyond that, you need to be able to monitor, detect, and respond in real time. Prevention is not the leading priority. It’s all about rapid response. Source photos by Aaron Wojack and Ulysses Ortega from the New York Times article.
-
Brian Contos shared thisFifty credible SOC playbooks in an afternoon. That part is real. The part nobody puts on the slide is that the playbooks were never what made a SOC mature. Anton Chuvakin published a piece this week revisiting an argument he made back in 2016, that you can jump technologies but you cannot jump maturity. Ten years and one agentic wave later he split the question in a way I have not seen anyone else do cleanly. Ask this of any capability you are about to buy. Does it exist because somebody had to spend time, or because somebody had to learn something about us? Time cost is jumpable. Playbook authoring, parsers, baseline detections, metrics plumbing, the paperwork of a mature program. Buy the jump. Take it. Learning cost is not. Knowing what your systems do. Knowing who owns them. Trusting your own telemetry enough to act on it. Earning the right to let a machine take an action without a human reading it first. Those are not documents. They are things an organization learned by operating. The line that landed hardest for me was about failure. A cargo cult SOC used to fail loudly and hilariously. Now it fails quietly. The agent produces a plausible investigation, the ticket closes, the metrics look fantastic, and nobody notices the context was invented and the containment hit the wrong host. I have been around long enough to remember SOAR selling the same promise. Every failed SOAR program I watched died on ownership and stamina, not on technology. Buy the jump on the paperwork. Climb the rest. Which of your mature looking artifacts were generated in the last ninety days, and has anyone checked whether they match how your team actually responds?
-
Brian Contos liked thisBrian Contos liked thisMany security teams are being asked to secure growing AI adoption without additional funding. For now, that means playing “prevent defense”—prioritizing controls that can stop the most damaging outcomes while the business determines how AI will deliver value. GuidePoint Security's Gary Brickhouse recently spoke with CRN about why greater clarity around AI use and ROI could make 2027 a turning point for AI security investment. Read more: https://okt.to/HIhEGFBigger AI Security Budgets Should Be Coming In 2027: GuidePoint CISOBigger AI Security Budgets Should Be Coming In 2027: GuidePoint CISO
-
Brian Contos liked thisBrian Contos liked thisHuman ingenuity never ceases to inspire. Truly humbled to be named to the TIME 100 Next list alongside incredible world leaders, activists, innovators, artists, athletes, and so many others pushing the boundaries of what’s possible. This recognition belongs to Cyera. Tamar Bar-Ilan, Yonatan Itai and I started it five years ago on the belief that data would be the foundation for securing the AI future. Our customers have always been the compass, and our incredible team has made it all a reality. Thank you to every person who has been on this journey with us and to Tharin Pillay for the thoughtful conversation. https://lnkd.in/grAp6_b8
-
Brian Contos liked thisBrian Contos liked thisApparently the best way to get me to travel halfway around the world for a cybersecurity conference is to host it somewhere with really good Gujarati food. 😂🇮🇳 Security BSides Ahmedabad was a special one. Getting to be back in Gujarat, on stage talking about security, surrounded by such an energetic community felt like a pretty great full-circle moment. Great conversations with customers and prospects, catching up with old friends, making new ones, and spending quality time with the Upwind Security crew. Huge thank you to Nikhil S. and the entire Security BSides Ahmedabad team for the incredible hospitality and for building such a strong security community. And Simarpreet Singh, Diksha Gupta, Himanshu Chaudhary Hui Juan Lee + the Upwind Security team, it was so great spending time with you all and meeting customers and eating a lot of food while calling it all of it work. 😂 Cloud security. Customers. Community. Chai. ☕ #BSidesAhmedabad #Upwind #Cybersecurity #CloudSecurity #SecurityCommunity #India
-
Brian Contos liked thisBrian Contos liked this* Steam-powered AI – or how history goes around in circles * Why is it that AI brings efficiency and productivity gains, yet its overall resource consumption just keeps on growing? For example, we recently found a bug in some of our code – routine stuff, really. With the help of AI tools, the bug got fixed in two days. The developers estimate that without those tools it would have taken two weeks – seven times longer. What an amazing saving! So, we could have just switched off the computers and turned out the lights for 12 days since the work was done, right? Needless to say, nobody switched anything off! Everyone involved always has other things to do: we’ve got a big, demanding operation to run. More bugs will turn up, and if they don’t we can always crank out some new features – with help from that same AI, too. All this is reminiscent of the Jevons paradox from the 19th century: the arrival of new, advanced steam engines made it possible to do a whole lot more while using a lot less coal. In other words, savings in resource (coal) usage per application were huge – and plain to see. But the result was paradoxical: demand for coal didn’t fall; steam engines started to be installed everywhere and so demand for coal skyrocketed. It’s the same today. If AI makes certain tasks quick and easy (when they used to be long and difficult), we start doing them thousands of times more often. And the data centers where all this so-called AI runs will be guzzling ever more gargantuan amounts of electricity. Basically, savings and efficiency gains mean being able to do the same amount of work while consuming less resources. Progress, though, usually works the other way round: you do A WHOLE HECK OF A LOT MORE and so in the end consume more – cancelling out the savings, and then some. Not a very pleasant thought as regards the environmental activism, but so far, that’s pretty much how scientific and technological revolutions have worked.
Experience & Education
Licenses & Certifications
-
Certified Information Systems Security Professional (CISSP)
CISSP and CompTIA Security+ Network+ A+ Server+ Instructor
Issued
Volunteer Experience
Publications
-
I Spy (With My Five Eyes)
Documentary
See publicationThis interactive documentary details how the Five Eyes spy network (intelligence agencies of USA, Canada, UK, Australia and NZ) sought a new justification for its existence after the USSR's collapse and found it in digital communications.
2018 Canadian Screen Awards
Nominated for Best Original Interactive Production for Digital Media -
Physical & Logical Security Convergence
Elsevier
This book begins with an overall explanation of information security and physical security and why approaching these two types of security in one way, called convergence, is critical in today's changing security landscape. It then details SEIM-related incident detection and response capabilities. This is followed by detailed implementation examples, taking the reader through cases addressing various physical security technologies such as video surveillance, HVAC, RFID, access controls…
This book begins with an overall explanation of information security and physical security and why approaching these two types of security in one way, called convergence, is critical in today's changing security landscape. It then details SEIM-related incident detection and response capabilities. This is followed by detailed implementation examples, taking the reader through cases addressing various physical security technologies such as video surveillance, HVAC, RFID, access controls, biometrics, and more.
Other authorsSee publication -
Enemy at the Water Cooler
Syngress
See publicationThe book covers a decade of work with some of the largest commercial and government agencies worldwide in addressing cybersecurity related to malicious insiders. It explores organized crime, terrorist threats, and hackers. It addresses the steps organizations must take to address insider threats at a people, process, and technology level.
Honors & Awards
-
Best Cybersecurity Podcast
Cybersecurity Insiders - Excellence Awards
The IoT Security Podcast https://cybersecurity-excellence-awards.com/candidates/the-iot-security-podcast/
-
Best Cybersecurity Podcast
Cybersecurity Insiders - Excellence Awards
The Cybersecurity Excellence Awards is an annual competition honoring individuals and companies that demonstrate excellence, innovation and leadership in information security. https://cybersecurity-excellence-awards.com/
-
I Spy (With My Five Eyes) Nominated for Best Original Interactive Production for Digital Media
Canadian Screen Awards
This interactive documentary details how the Five Eyes spy network (intelligence agencies of USA, Canada, UK, Australia and NZ) sought a new justification for its existence after the USSR's collapse and found it in digital communications.
Languages
-
English
Native or bilingual proficiency
Organizations
-
Forbes Technology Council
Contributor
- PresentSharing insights on asset intelligence with the Forbes Technology Council, an invitation-only organization for successful senior-level technology executives.
-
InfraGard
Member
- PresentInfraGard is a partnership between the FBI and members of the private sector. The InfraGard program provides a vehicle for seamless public-private collaboration with government that expedites the timely exchange of information and promotes mutual learning opportunities relevant to the protection of Critical Infrastructure. With thousands of vetted members nationally, InfraGard's membership includes business executives, entrepreneurs, military and government officials, computer professionals…
InfraGard is a partnership between the FBI and members of the private sector. The InfraGard program provides a vehicle for seamless public-private collaboration with government that expedites the timely exchange of information and promotes mutual learning opportunities relevant to the protection of Critical Infrastructure. With thousands of vetted members nationally, InfraGard's membership includes business executives, entrepreneurs, military and government officials, computer professionals, academia and state and local law enforcement; each dedicated to contributing industry specific insight and advancing national security.
-
Ponemon Institute
Distinguished Fellow
- PresentPonemon Institute was founded in 2002 by Dr. Larry Ponemon. Headquartered in Michigan, Ponemon Institute is considered the pre-eminent research center dedicated to privacy, data protection and information security policy. Our annual consumer studies on privacy trust are widely quoted in the media and our research quantifying the cost of a data breach has become valuable to organizations seeking to understand the business impact of lost or stolen data. Our research has made Ponemon Institute…
Ponemon Institute was founded in 2002 by Dr. Larry Ponemon. Headquartered in Michigan, Ponemon Institute is considered the pre-eminent research center dedicated to privacy, data protection and information security policy. Our annual consumer studies on privacy trust are widely quoted in the media and our research quantifying the cost of a data breach has become valuable to organizations seeking to understand the business impact of lost or stolen data. Our research has made Ponemon Institute a thought leader on trends in corporate and government practices, consumer perceptions and potential cyber and insider threats that will affect the collection, management, and safeguarding of sensitive and confidential information. Our benchmark studies on privacy and data security practices provide guidance on how organizations can become better at building trust with their customers, employees, and business partners. We offer independent assurance and strategic advisory services to help companies enhance the effectiveness of their privacy and data protection programs and meet compliance and regulatory requirements in the United States and abroad. Ponemon Institute is the parent organization of the Responsible Information Management Council. The RIM Council has members from leading multinational corporations and focuses on the development and execution of ethical principles for the collection and use of personal, sensitive, or confidential information. The RIM Council enables both seasoned privacy practitioners and those new to the profession to network and learn how to be as successful as possible in their organizations in both the private and public sectors.
Recommendations received
32 people have recommended Brian
Join now to viewOther similar profiles
Explore more posts
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content