The fix had been sitting there for twenty-one years when Jack Cable and I recorded this.
“If you take SQL injections, where we’ve had parametrized queries in MySQL since 2004, so 21 years now, and yet we still see these SQL injections coming up, despite having been fully preventable for decades.”
Jack Cable, at 23:59
Jack Cable is the co-founder and chief executive of Corridor, and since October 2026 the head of security products at Cognition, which bought it. He found his first vulnerability at 15, won the Air Force’s bug bounty challenge in high school, spent the summer after it at the Pentagon’s Defense Digital Service, helped scan the election infrastructure of all 50 states from inside CISA in 2020, worked on open source security legislation as a Senate fellow, and spent two years at CISA leading on Secure by Design, including a pledge that more than 250 software makers signed. He started Corridor in January 2025 with Ashwin Ramaswami, his partner on the election work, on one idea: the vulnerabilities attackers exploit are mostly the ones the industry has known how to prevent for decades, and AI can finally help prevent them where the code gets written. We recorded in 2025, after Corridor’s seed round, and published the episode on December 16, 2025, with 45 words of show notes. This page is the rest of it: the negative dollar, Hack the Air Force, a startup inside government, the pledge, the bugs that keep getting exploited, why SQL injection is still with us, scanners that don’t know your code, the pull requests nobody can review, the co-founder you already trust, a first round in March and a seed in April, knowing which bugs to fix, proving it, why defenders should get the tools, software that isn’t written by humans, the summer Jack picked the Pentagon, falling in love with a problem, and what Corridor did in the year after the episode came out. There are four things to try along the way: a quiz, a scorecard for whether the old bug classes are gone from your code, a calculator for how much of your scanner’s output is noise, and a timeline you can filter.
Negative one dollar
I asked whether all of this made him a hacker, and Jack told me how he became one, by accident.
“And as a curious kid, I entered zero into the transaction amount and noticed that the payment went through, and I thought, huh, that’s a little weird. And then I entered negative $1. And to my surprise, that also went through and actually took a dollar from that account.”
Jack Cable, at 3:17
He was 15 and building an integration with the API of “a cryptocurrency website, of all things.” Two years later he told Defense One which one it was, a Bitcoin tipping service called ChangeTip that has since shut down, and that the same flaw let him send negative amounts to other users, which moved their money into his account (Defense One). My follow-up was the one everybody asks, whether the real decision was how many zeros to add to the next one, and Jack’s answer was that there wasn’t one: “Yeah, exactly. I didn’t even really consider the alternative.” The company had a bug bounty. They paid him, he thinks, $1,000, “which at the time, as a 15-year-old, was a lot, still is a lot of money.”
He also credits the timing. “I got into security at a time where really vulnerability disclosure, bug bounties were much more the norm,” he said, and his record shows how much he made of it. From November 2015 to December 2020 he hunted on HackerOne, where he reported more than 350 valid vulnerabilities to organizations including Google, Yahoo, Facebook, Uber, Salesforce and the Department of Defense, and ranked in the top 100 on both HackerOne and Cobalt. From 2018 he was also on the Synack Red Team (LinkedIn).
The habit he describes from those years runs through everything else on this page, from bug bounties to the policy work. “It comes down to trying lots of different approaches, seeing what works. Many don’t. But often there is a way to have an outsized impact, whether that’s with a vulnerability, whether that’s through the Secure by Design work, for instance.”
Hack the Air Force, then the Pentagon
Jack’s own short version came at the top of the episode: “I got into government work originally through the Hack the Pentagon bug bounties. So when I was in high school, I won the Hack the Air Force competition and then went to work for the Pentagon after high school.”
The long version is in the Air Force’s numbers. Hack the Air Force ran from May 30 to June 23, 2017. More than 270 vetted researchers took part, they found 207 valid vulnerabilities and earned more than $130,000, and the Air Force singled out one participant, a 17-year-old who submitted 30 valid reports and earned the largest bounty total of the challenge (U.S. Air Force). That was Jack. Defense One counted more than 30 of the 207 as his, and reported that the run lifted him to fifth in HackerOne’s third-quarter rankings (Defense One). In September he posted the souvenir.
Completed federal bug bounty trio with Hack the Air Force coin. Props to @Hacker0x01 + @DefenseDigital for organizing such awesome programs.
Jack Cable (@jackhcable) · September 13, 2017
The Pentagon part came the next summer. From June to September 2018 he worked full time at the Defense Digital Service, the Pentagon’s in-house team of engineers, and advised it until 2022, helping lead the Hack the Marine Corps bug bounty and advising the Hack the Pentagon program (LinkedIn). That December, TIME put him on its list of the 25 most influential teens of 2018, at 18 (TIME). By then he was at Stanford. In January of his first year he helped the university’s information security office launch a bug bounty for students and staff, which Stanford said put it among very few universities with one (Stanford University IT), and from 2019 he was a research assistant at the Stanford Internet Observatory and head teaching assistant for its Hack Lab course (LinkedIn). He finished a bachelor’s in computer science there (TechCongress). How he chose the Pentagon over a Silicon Valley internship that first summer is near the end of the episode, in the advice to his younger self.
A startup inside government
When I asked about his co-founder, Jack went back to 2020.
“We built our own open source attack surface management tool called Crossfeed, and we used that to secure election infrastructure. So we rolled that out to scan all 50 states, every county in the US, to look for vulnerabilities in their internet-facing sites.”
Jack Cable, at 10:21
His LinkedIn has the count. As an election security technical advisor at CISA from June 2020 to January 2021, he led the agency’s first passive, opt-out vulnerability scanning program, built on Crossfeed, which assessed all 50 states and over 2,500 counties before the 2020 election (LinkedIn). Crossfeed’s code is public, a joint project of CISA and the Defense Digital Service that continuously maps an organization’s internet-facing attack surface (GitHub). Through it, he said, they “reported hundreds of [high and] critical vulnerabilities to states,” and the part he remembers best is what it felt like to build: “we had essentially almost built a startup within government.”
Between government jobs he kept finding ways to hand defenders an edge. In April 2021, while still at Stanford, he noticed that the operators of the QLocker ransomware checked payments so loosely that changing one letter of a transaction ID from lowercase to uppercase unlocked a victim’s files, and he used it to recover data for 50 victims, about $27,000 in ransom they didn’t have to pay, before the gang fixed it (CyberScoop). The Secretary of Homeland Security noticed.
Great work by @jackhcable! From disrupting #ransomware schemes to working with @CISAgov to #Protect2020, you are a tremendous example of how even a single person can make a difference.
Secretary Alejandro Mayorkas (@SecMayorkas) · April 23, 2021
That July he launched Ransomwhere, a crowdsourced tracker of ransom payments, which had counted more than $32 million in 2021 payments by launch day (TechCrunch). From May 2021 to February 2022 he was a security architect at Krebs Stamos Group, the consultancy Chris Krebs and Alex Stamos founded that year, leading purple team engagements and incident response (LinkedIn, The Record). Stamos comes back into this story twice.
Secure by Design, and a pledge
In 2022 Jack spent most of a year in the Senate as a TechCongress fellow for the Homeland Security and Governmental Affairs Committee under its chairman, Gary Peters, working on open source and election security, including the bipartisan Securing Open Source Software Act (LinkedIn). Peters and Rob Portman introduced the bill on September 22, 2022, after their hearing on the Log4j vulnerability, to have CISA build a framework for the risk in the open source code the federal government runs (Senate Homeland Security Committee).
In January 2023 he went back to CISA as a senior technical advisor, leading on its Secure by Design and open source software security work. That April, CISA and its international partners published the principles behind the campaign, a guide to shifting the burden of security from customers to the companies that make software (CISA). The pledge followed at RSA Conference in May 2024. Sixty-eight companies, including AWS, Microsoft, Google, Cisco and IBM, committed to make a good-faith effort within a year toward seven goals, among them more multi-factor authentication, fewer default passwords, reducing entire classes of vulnerability and publishing a vulnerability disclosure policy (The Register).
On the tape it’s one sentence: “Started a pledge and got commitments from now over 300 companies, and there’s been some good tangible results coming out of that.” When he left CISA in January 2025, the number he gave CyberScoop was more than 250, and he recalled that early feedback from industry was that the pledge wouldn’t get traction (CyberScoop). His LinkedIn says 250 of the world’s largest technology manufacturers, and CISA’s own page now says hundreds (CISA). The same exit interview has the sentence that became Corridor: attackers keep getting in through relatively simple vulnerabilities, and software makers “have also known how to prevent these at scale for decades.”
The bugs that keep getting exploited
I asked what he saw that made him start Corridor, and his answer started at CISA.
“And I’d say one observation was that we would talk about things like rooting out entire classes of vulnerabilities from your products, with the recognition that at the end of the day, many of the vulnerabilities that we see out there exploited aren’t all that complicated. They are preventable.”
Jack Cable, at 6:21
He named them: “Things like SQL injections, cross-site scriptings, but even vulnerabilities like authorization flaws, where we really ought to be doing more to prevent these at the source.” Later in the episode he added the ones in the government’s advisories about nation states: “often it is SQL injections, command injections, memory safety vulnerabilities, things that if we were talking, I don’t know, ten years ago, we would think wouldn’t be a problem in the future.”
The public record keeps the same list. In March 2024, CISA and the FBI published a Secure by Design alert on SQL injection, after a campaign that exploited it in a managed file transfer product and hit thousands of organizations. It opens from the same place Jack does, “Despite widespread knowledge and documentation of SQLi vulnerabilities over the past two decades,” and notes that others have called SQL injection an “unforgivable” vulnerability since at least 2007 (CISA and FBI). The product was Progress Software’s MOVEit Transfer. CISA’s advisory says the Cl0p ransomware gang began exploiting a previously unknown SQL injection flaw in it, CVE-2023-34362, on May 27, 2023 (CISA), and by June 2024 Emsisoft’s running count of the victims stood at 2,773 organizations and more than 95 million people (Emsisoft). The OWASP Top 10 for 2025 still has broken access control at number one and injection at number five (OWASP).
Authorization flaws are the class Jack kept coming back to, because they’re the hardest to find without context. A scanner can match a dangerous string pattern. It can’t know who is supposed to see a record unless it understands what the application is for, which is where the next part of the conversation went.
Fully preventable for decades
Jack’s date checks out. MySQL 4.1, the release that added prepared statements, went to beta in June 2004 and to production that October (Wikipedia). Prepared statements are the fix the CISA and FBI alert prescribes, word for word: “developers should use parameterized queries with prepared statements to separate SQL code from user-supplied data” (CISA and FBI). The query and the user’s input travel separately, so the input can never be read as code. It’s a one-line habit, it’s been in the major databases for two decades, and SQL injection still makes the news, as MOVEit did.
What Jack took from that is optimism, which is not what you’d expect from someone who spent two years asking companies to fix the same bugs. “So it is possible to have a future where code is more secure by design,” he said, and he gave an example of a company that already lives there: “Google, for instance, has talked about how they haven’t had SQL injections or cross-site scripting vulnerabilities in their products.”
The record makes his point better, with one careful word. Google’s Christoph Kern writes that the company’s safe coding practices “have nearly eliminated vulnerabilities such as Cross-Site Scripting (XSS) and SQL injection” in the code built on them (Google Research), and his 2024 article in ACM Queue reports “near-zero residual rates” of common defect classes across hundreds of applications built by thousands of developers (Google Research). Nearly, not never. The method is the part worth copying. Google didn’t train every developer to remember the rule. It made the safe way the default way, in the frameworks and libraries its developers already use, so that writing the bug takes extra effort.
That is also his answer to the most common objection in security, that nothing is ever fully secure.
“But I think it’s really about kind of raising the cost on adversaries. And yes, we’re not going to ever be able to root out every single vulnerability from a piece of software, but I don’t think you can say rationally that we’re always going to face memory safety vulnerabilities or we’re always going to have to deal with SQL injections, because we know how to prevent those, because companies have already been able to do so.”
Jack Cable, at 26:47
Scanners that don’t know your code
The second half of Jack’s founding observation was about the tools. He had used them from both sides, as a bug hunter who reads code and as the person at CISA asking companies why the same bugs kept shipping.
“And if I think about, for instance, when I do a bug bounty or a pen test and I’m kind of researching for vulnerabilities, if I have access to the code, searching through for vulnerabilities, that looks very different than the existing approach these tools are taking, which essentially is just kind of searching through with kind of static, pre-canned templates, because it doesn’t actually understand how the code base functions.”
Jack Cable, at 6:21
His alternative was to make the scan read the code the way a researcher does: “it’s now much more possible to have tailored scans that actually take into account that functionality, that says, okay, this is a vulnerability for these reasons. We can see the path by which that’s exploitable, or vice versa, this one we really don’t need to care about for other reasons.” That last clause matters as much as the first. A tool that can explain why a finding doesn’t matter saves more engineering time than one that finds a little more. Developers say the same thing in their own words. On the Hacker News launch of another AI code scanner in July 2025, one wrote:
"cookie cutter checks aren't that effective for catching anything but the most basic vulnerabilities."
dnsbty, on Hacker News, July 2025
When Axios covered Corridor’s seed round that August, it described the target as context-heavy issues, such as authorization flaws, that traditional tools often miss (Axios). By 2026 the company was doing the indexing ahead of time, so that a coding agent could get guidance specific to the codebase while it worked.
Nobody can review every pull request
The episode’s title comes from this exchange, and the line that started it was mine.
“Love it. I wish I could review every pull request, just can’t.” Jack had just said the same thing from the vendor’s side: “for most companies, it’s not possible to have a security person go and review every pull request, every piece of code that’s coming in. But we can get closer, we can do better now with AI and really augment and increase the work that these security professionals are able to do.” Past a certain size, security review becomes something people ask for when they notice a change “maybe touches authentication,” which means it happens when someone remembers.
The volume is the part that changed. “Right now, developers can ship code many times faster than they were able to before,” Jack said. “The amount of code that’s getting written or generated is increasing significantly. And I just don’t think that the existing security tools are going to be able to keep up with that.” The week before Corridor announced its seed round, Veracode published its test of more than 100 language models on 80 coding tasks. The generated code introduced a security flaw in 45 percent of cases, the models failed to defend against cross-site scripting 86 percent of the time, and larger models didn’t do meaningfully better than smaller ones (Veracode).
His other condition is the one most security tools fail. “So working towards this future where security can go really hand in hand, where it enables development and doesn’t slow things down.” My version of the same idea is that adoption is the test. A security tool that developers route around has a score of zero, however good its findings are.
You’re marrying the person
The co-founder question has one answer Jack would give anyone. His co-founder is Ashwin Ramaswami, whom he met as an undergraduate at Stanford, worked with at the Stanford Internet Observatory, and then built Crossfeed with at CISA. “So, when we started talking again this year, when we started thinking about, okay, how can we really drive forward secure by design, he was the first person I thought of, because we had done it before. We knew we worked well together.” The general version: “it’s really beneficial to have someone who you already have that trust with. Because ultimately it’s a many years long commitment.”
My version was shorter. “You’re marrying the person, basically, professional sense, right? You’re gonna go through a lot together. There’s gonna be, like, very high highs, and there’s gonna be very low lows.”
Between the election work and Corridor, Ashwin had left CISA in 2023 and run for the Georgia State Senate in 2024. Forbes reports that the call came the day after he lost, in November 2024, and that the two of them launched Corridor about two months later (Forbes). Ashwin became chief technology officer (Axios). Both are engineers, so the CEO question had no obvious answer. “As it turned out, I became the CEO. So nowadays, you know, I still try to code when I can, but a lot more of my time is focused on all the other parts of running a business.” He thinks the years at CISA, spent making the case for Secure by Design to companies that didn’t have to listen, trained him for the part of the job that is telling the story: “Because I think that’s so much of what a company comes down to, is the brand. It’s how people perceive you.”
Angels in March, a seed in April
The funding came fast, and from people who had built what Corridor wanted to be.
“And for context, we started this in January of this year. We did an angel round in March from folks like Alex Stamos, Christina Cacioppo, CEO of Vanta, a number of other security founders, Feross from Socket, Dylan from TruffleHog, and a bunch of CISOs as well. And then we followed that up in April with a seed round that was led by Conviction, Sarah Guo.”
Jack Cable, at 13:43
Two of those names have been on this show. Dylan Ayrey of Truffle Security was episode 9, in 2021, and Feross Aboukhadijeh of Socket was episode 36, in 2024. Vanta is personal for Jack: “Christina, for instance, from Vanta, I was the first intern at Vanta in 2019, back when there were six people.” That’s his account. It isn’t on his LinkedIn, and I haven’t seen it anywhere else.
When the seed was announced in August 2025, Axios put it at $5.4 million, led by Conviction, with angels including Alex Stamos, Bugcrowd founder Casey Ellis and Duo Security co-founder Jon Oberheide, and with Stamos joining as chief security officer. Stamos, who has run security at Yahoo and Facebook and was then chief security officer at SentinelOne, told Axios that few of the Stanford students he’d met were as dedicated to security as these two. Corridor had five employees, and its customers included Cursor, Mercury and GreyNoise (Axios). Corridor’s own post that day, six months after it started, named the problem in nine words: “The same, basic, preventable vulnerabilities continue to be exploited” (Corridor).
Knowing which ones to fix
The best story in the episode is about being wrong. Jack’s rule for a new company is to “throw all of your assumptions aside and see what actually works in practice,” and he gave Corridor’s first big assumption as the example.
“But through the course of talking to folks, I’d say a lot of what we realized is that, okay, the problem isn’t just fixing flaws, it’s knowing which ones to fix. If you have thousands and thousands of vulnerabilities, and by the way, if 90-plus percent of those are false positives, then your biggest problem isn’t, once you’ve identified this is a valid vulnerability, going and fixing it. It’s knowing which ones are valid.”
Jack Cable, at 14:49
The 90 percent is his estimate, not a study, and anyone who has run a static analyzer on a large codebase will recognize it. A security engineer described the experience on the same Hacker News thread, after a customer insisted on running a popular scanner on the company’s 20 million lines of code:
"Two of us spent two weeks clearing false positives. Absolutely nothing was left."
wglb, on Hacker News, July 2025
Neatsun Ziv made the same point on episode 91, and our post on why shift-left security keeps failing walks the public numbers on how few findings are ever exploited. Jack’s version adds the step that comes first, which is validation. Before anyone can prioritize by exploitability, someone has to find out which findings are real, and that is the work that eats the security team’s week. His filter for what Corridor builds next came from the same lesson: listen to every customer request, then ask which “are common problems, are going to be things that we can go and scale out to the next 50 or 500 companies.”
Proving it
His proudest moment so far was the first sign the approach worked: “we’re discovering vulnerabilities that they previously hadn’t known about, that, like, we’ve heard are kind of on par with what they’d expect to see, for instance, from a human attacker in their bug bounty.” For a former bug hunter, that’s the comparison that counts.
The balance he described is the part I’d quote to anyone buying an AI security product. “Let’s face it, there’s a lot of AI products, there’s a lot of things that say they do something, and, you know, there are kind of promises that are made but not kept.” His answer is to use the models where they’re strong and not elsewhere: “not using it as a panacea, but rather kind of taking the benefits and, where necessary, for instance, combining them with more deterministic approaches in order to really build the best products.”
The hardest moments were the early sales conversations, when “people would say, okay, but how is this different from X tool?” What got him through it was a definition of the company that doesn’t depend on this month’s product.
“It’s knowing that our company is not our current product. Today, that will evolve. It’s really people, it’s our mindset that makes us unique, makes us different. And the fact that we can adapt, and especially given this age of AI, where every month or two we get a new model that’s going to go and outperform all the previous models, you can’t tie yourself to any one approach.”
Jack Cable, at 19:33
Better to give defenders more tools
I asked the uncomfortable question. A tool that finds fresh vulnerabilities in code works for whoever runs it, and attackers have code too. Jack had an answer ready, because the security community has had this argument before, about vulnerability disclosure and dual-use tools like Metasploit, the exploitation framework H. D. Moore created in 2003 (Wikipedia).
“And really, my perspective there is that it’s always better to give defenders more tools. The attackers already have no trouble going, finding, exploiting vulnerabilities. And, right, you need to, for every one vulnerability attackers are exploiting, you need to make sure that 100 others are prevented to avoid them getting in.”
Jack Cable, at 21:56
The asymmetry is the whole argument. An attacker needs one bug that works. A defender needs the other hundred closed, which is a volume problem, and volume is what automation is for. Then he tied it back to Secure by Design, because finding bugs faster is only half the prize: “if defenders can not just find vulnerabilities but learn from their mistakes, look at the common root causes and see how to prevent those from happening in the first place, then we can start to move towards fundamentally more secure software.” My addition was about the rules. Defenders follow them and attackers don’t, “So you have to, like, level the playing field a little bit by building all of these great, amazing tools, just like Corridor.”
Software that isn’t written by humans
When I asked him to look ahead, he started from where software is going.
“Maybe it’s helpful to kind of think about where software will be in the future, where I think we will move to a point where the vast, vast majority of software isn’t written by humans. It’s generated by AI, likely with the supervision and help of humans.”
Jack Cable, at 23:59
His forecast for what that does to security was the opposite of the usual one. “But let’s also ground it, because there’s a lot of people who talk about kind of all the unknowns that are supposedly associated with AI.” History says the old bugs stay unless something changes, so the risk he worries about is scale: “Because if we’re only going to, I don’t know, 10x, 100x the amount of code out there, I don’t think we want a future where that code is riddled with the same exact types of vulnerabilities as we’ve been struggling with today.”
He also said where the fix won’t come from, and it’s the sentence I’d most like to check again in a few years.
“I don’t think that will be solved at the model level, because so often security is contextual. It’s not about just, like, a piece of code in isolation, but seeing how it interacts with a broader code base, especially when it comes to business logic or authorization issues, which we found that AI can do quite well in understanding how applications actually work.”
Jack Cable, at 23:59
My summary on the tape was a plea not to automate the history: “let’s not train on the mistakes of the past only.” A year later, Corridor’s own work read like a test of his sentence. In October 2026 it released a small model of its own, and the job it gave the model was choosing the codebase-specific guidance, which is to say, the context. That’s below.
Follow the direction of your heart
I asked about the legacy he wants, and the answer explains the career. “A recognition that it’s possible to scale impact beyond just one person.” He liked bug bounties, and still came to feel there was “probably a better way than kind of doing these one-off tests.” The question became scale. “How can I start to, for instance, through policy, scale and reach thousands, tens of thousands of companies?” The goal is to look back and say “I actually helped secure the critical software that we all rely on.”
Then I asked what he’d tell his younger self, and he told the story of the summer after high school.
“What I would tell myself is, I think, really to kind of always just follow the direction of your heart. And just to give one example, where when I was 18 or 17, I was deciding, okay, what to do the summer after high school, and had the standard tech internship at a company out in SF, and was weighing that against working at the Pentagon.”
Jack Cable, at 31:29
He chose the Pentagon, and it set the next seven years: “that was what started my interest in policy and working in government.” The principle he drew from it is a good one for anyone choosing between two offers: “But I think a lot of it comes down to thinking about orthogonal directions. What is going to be the most unique experience? What sets you apart from everyone else?” The prestige was on the other side. “So I think it ultimately is around taking the more interesting path, and you never know what’s going to come out of that.”
Fall in love with a problem
We have a lot of founders listening, so I asked him for a problem he’d pay someone to solve. He gave a method instead, which is better. “I think often the biggest thing to build is just your own intuition, your own sense of direction, and having been in the seat of your customer yourself.” Then he made the case for starting now: “kind of in this age of AI, where the barriers to build a product are a lot lower, the kind of range of what you can do with a product, the possibilities are more unbounded, I think that there’s kind of no better time to be starting something.” He finished in eleven words: “Fall in love with a problem. Go experience the pain firsthand.”
It’s also his own biography. He spent years on the inside of the problem, as a bug hunter, a Pentagon advisor and a technical lead at CISA, before he built a company around it. My addition from the founder’s side was about the first version of anything: “You gotta revisit the problem, revisit the pain, and listen to the market.”
What happened after the tape
Corridor’s public record from the summer of 2025 to the end of October 2026, in order.
- August 5, 2025: the seed, and Alex Stamos. Corridor announced its $5.4 million seed round, led by Conviction, and Stamos joined as chief security officer (Axios).
- December 3, 2025: a category name. “Introducing Agentic Coding Security Management,” the company’s term for security built into AI coding workflows (Corridor). Around it came a Factory integration in October, a listing on the AWS Marketplace and support for Cursor’s hooks in December (Corridor).
- December 16, 2025: this episode.
- March 18, 2026: a $25 million Series A. Led by Felicis, with Conviction, Lux Capital, Datadog and others, at a $200 million valuation, with Sublime Security and Pylon named as customers (Forbes, EZ Newswire). Jack’s line in the release: “Secure by design can’t just be a slogan.”
- May 2026: plugins. A Cursor plugin and a Codex integration (Corridor).
- June 4, 2026: Congress. Jack testified before the House Homeland Security subcommittee on cybersecurity and infrastructure protection, next to witnesses from Google Threat Intelligence, the Frontier Model Forum and the Electronic Frontier Foundation (House Homeland Security Committee). His testimony could have been read off this tape. “Attackers generally aren’t exploiting new kinds of vulnerabilities,” he said, and he told the committee that across Corridor’s customers, giving the coding agent the right context at the planning stage reduces vulnerabilities by 60 percent, by the company’s measure (Tech Policy Press).
- July 2026: the bugpocalypse. A talk, published by AI Engineer, on what it means that AI models now find serious vulnerabilities at scale, and the summer’s launches, among them pre-commit scanning for Cursor’s cloud agents, a merge policy security teams can enforce, and Corridor Agent for triage and fixes (Corridor).
- August 14, 2026: the coding agent as the target. Corridor’s research showed coding agents could be tricked into running malware through an ordinary support ticket.
New research from @corridor: we found that coding agents - even with models like Fable - can be trivially tricked into running malware.
Jack Cable (@jackhcable) · August 14, 2026
We connected coding agents to our support system, filed a ticket, and got them to exfil secrets and run malware.
- September 29, 2026: Stamos moves to Cognition, the company behind the coding agent Devin, as its security chief (Let’s Data Science).
- October 2026: a model for the context. On October 7, Jack shared a Washington Post story on agents, prompted with ordinary data-gathering tasks, that attempted rudimentary hacks against government websites when they hit roadblocks, research Corridor did with Transluce (LinkedIn). Two weeks later Corridor released Argos.
Incredible work from the Corridor team!
Jack Cable · October 21, 2026
Corridor: Today we’re introducing Argos, Corridor’s first security model, to steer coding agents toward writing more secure code.
Argos reads a coding agent’s plan and selects codebase-specific security guidance before feature implementation. These requirements become part of the plan the agent uses to write the code, which means more secure code by default, and less vulnerable code for developers and security teams to fix later.
- October 28, 2026: Corridor joins Cognition. Jack and Ashwin announced it on Cognition’s blog, where the team will work on Devin Security (Cognition). Jack’s LinkedIn now says head of security products at Cognition, and Ashwin’s says member of technical staff.
Today, Cognition is acquiring @corridor.
Cognition (@cognition) · October 29, 2026
Securing the world’s software is the most important problem in technology today.
Let’s get to work!
Today, Corridor is joining Cognition! Together, we’ll be taking Devin Security to the next level.
Jack Cable · October 29, 2026
The next 6-12 months represent the most consequential period in the history of cybersecurity. AI agents can now run end-to-end attacks, mostly by exploiting vulnerabilities we’ve known about for decades.
Fortunately, defenders can use the same capabilities to root out exploitable vulnerabilities before attackers find them. For the first time, it’s possible to not just patch vulnerabilities one at a time, but systematically eliminate them at the source, making Secure by Design a reality.
Fixing vulnerabilities is fundamentally a software engineering problem, which is why we’re so excited to join forces with Cognition.
The past two years have been a whirlwind, and I’m incredibly grateful to our team, our customers, and our investors for the support along the way. I can’t wait to show you what we build next.
Read in order, the record is a person saying the same thing for ten years with better tools each time. At 15 it was a bug report. At CISA it was a pledge. At Corridor it was a product that hands the agent the right context before it writes the code. At Cognition it’s Devin Security, inside the company that makes the coding agent. Corridor’s homepage puts it in six words: “Security for Software That Builds Itself” (Corridor).
Did you catch it? A six-question quiz
Interactive · six questions
Did you catch it?
Pick an answer for each, then check. Every answer is on the tape.
-
What did Jack type into the transaction amount before he tried negative one dollar?
- Zero
- One million
- A letter
a. "I entered zero into the transaction amount and noticed that the payment went through, and I thought, huh, that's a little weird." Transcript, 3:17.
-
About how much did the company pay him for that first bug?
- $100
- $1,000
- $10,000
b. "They paid me, I think, $1,000 for that first vulnerability, which at the time, as a 15-year-old, was a lot, still is a lot of money." Transcript, 4:20.
-
Which competition did Jack win in high school?
- Hack the Army
- The DEF CON capture the flag
- Hack the Air Force
c. "So when I was in high school, I won the Hack the Air Force competition and then went to work for the Pentagon after high school." Transcript, 1:46.
-
Corridor started out focused on fixing vulnerabilities. What did customers teach it the bigger problem was?
- Knowing which ones to fix
- Getting developers to read the reports
- Passing compliance audits
a. "the problem isn't just fixing flaws, it's knowing which ones to fix." Transcript, 14:49.
-
Since when, by Jack's count, has MySQL had parameterized queries?
- 1995
- 2004
- 2014
b. "where we've had parametrized queries in MySQL since 2004, so 21 years now" Transcript, 23:59.
-
What did Jack weigh against the Pentagon the summer after high school?
- Starting a company
- A year of travel
- A standard tech internship in San Francisco
c. "and had the standard tech internship at a company out in SF, and was weighing that against working at the Pentagon." Transcript, 31:29.
Are the old bugs gone from your code?
Jack’s claim is checkable for any codebase: the bug classes that keep getting exploited are the ones we know how to prevent, so the question is whether your code prevents them by default. This scorecard asks eight questions about the code your team ships and tells you where you stand.
Scorecard · eight questions
Are the old bugs gone from your code?
Answer for the code your team ships today, including what your coding agents write. If two answers fit, pick the lower one.
-
How does your code build database queries?
- Some queries are built by joining strings
- Mostly an ORM or parameterized queries, but nothing enforces it
- Parameterized by default, and a check blocks raw string queries
Prepared statements have been in MySQL since 2004. The section above has the record.
-
How is user input escaped when your pages and emails render it?
- Developers escape it by hand
- The framework escapes by default, with exceptions nobody tracks
- Escaping is the default everywhere, and every exception is reviewed
Google's near-zero rates for cross-site scripting came from frameworks that make the safe way the default.
-
How do your API routes decide who can see or change a record?
- Each handler checks, when someone remembers
- A shared helper that most routes use
- Deny by default, with an authorization test for every route
Broken access control is number one in the OWASP Top 10 for 2025, and it's the class Jack said scanners miss without context.
-
Does any code pass user input to a shell or a file path?
- Yes, and we haven't looked
- In a few places, behind wrappers
- No, only safe APIs that take arguments as data
Command injection was on Jack's list of what nation states exploit.
-
What language is your new code written in?
- New C or C++ that handles untrusted input
- Mostly memory-safe, with legacy code we haven't touched
- Memory-safe for new code, with a plan for the rest
Memory safety was the third class on his list, and the one with the clearest fix.
-
What do your AI coding agents know about your security rules before they write code?
- Nothing
- A rules file someone wrote once
- Codebase-specific guidance they get while planning, kept current
Jack told Congress that context at the planning stage cut vulnerabilities by 60 percent across Corridor's customers.
-
Who looks at security in your pull requests?
- Nobody, unless someone asks
- People, on the ones that get flagged
- An automated security review on every change, and people on the risky ones
Nobody can review every pull request by hand. That section is about what to do instead.
-
What happens to your scanners' findings?
- They pile up, and nobody triages
- Someone triages in bulk now and then
- Each one is validated, ranked by exploitability, and false positives feed back
"It's knowing which ones to fix." The calculator below puts a number on it.
The old bugs are probably still possible in your code, which is true of most startups and very fixable. Don't start with a scanner. Start with one class, usually SQL injection or authorization, make the safe way the default way in your framework, and add a check that fails the build when someone goes around it. Then do the next one.
You've made the safe way available, and now it needs to be the only way. Enforce the defaults with checks, give your coding agents the same rules before they write code, and put an automated security review on every pull request so your people only see the ones that matter.
The known classes are prevented at the source, enforced by checks, and taught to the agents, and your findings get validated instead of piling up. That's the pledge's third goal done in practice. Write it up so customers can check it, and keep measuring, because the code is growing faster than it used to.
How much of your scanner’s output is noise?
Jack’s 90 percent makes a quick estimate possible. This calculator takes the findings your scanners raise in a month, the share you expect to be false positives, and the time it takes to check one, and shows what the noise costs in a year.
Calculator
How much of your scanner's output is noise?
Your monthly findings, the share that turn out to be false positives, the minutes it takes to check one, and what an engineer's hour costs you.
With 400 findings a month and 90 percent false positives, 360 of them a month are noise and 40 are real. At 20 minutes each, checking the noise takes 1,440 hours a year, about $172,800 of engineering time spent proving that a finding doesn't matter.
The 90 percent default is Jack's estimate on the tape, "90-plus percent of those are false positives," not a measured rate, and yours depends on your tools and codebase. The 400 findings, 20 minutes and $120 an hour are placeholders, not research. The fix that shrinks every number at once is the one from the episode: prevent the known classes at the source, so there's less to find, and validate the rest automatically, so people only check the real ones.
From negative one dollar to Devin Security
Everything on this page sits on one timeline, from the first bounty to the acquisition. Filter it by thread.
Timeline · 2004 to 2026
From negative one dollar to Devin Security
Filter by thread. Dates are from the public record except where marked as Jack's own account on the tape.
- Prepared statements in MySQL
MySQL 4.1's production release (Wikipedia).
- Unforgivable
SQL injection is already called an unforgivable vulnerability, by the 2024 alert's account (CISA and FBI).
- Negative one dollar
At 15, on a Bitcoin tipping site, and a bounty of about $1,000.
- Hack the Air Force
More than 30 of its 207 valid reports, and the largest bounty total (U.S. Air Force).
- The Pentagon
The Defense Digital Service, instead of a tech internship (LinkedIn).
- TIME's influential teens
One of 25, at 18 (TIME).
- Stanford's bug bounty
Launched with his help in his first year (Stanford University IT).
- Vanta's first intern
When the company had six people.
- Crossfeed
CISA's opt-out scanning of election infrastructure, all 50 states (LinkedIn).
- QLocker
A one-letter trick that recovered files for 50 ransomware victims (CyberScoop).
- Ransomwhere
A crowdsourced tracker of ransom payments (TechCrunch).
- Dylan Ayrey on this show
Truffle Security, later one of Corridor's angels (episode 9).
- Securing Open Source Software Act
Introduced by Senators Peters and Portman (Senate Homeland Security Committee).
- Back to CISA
Senior technical advisor, Secure by Design and open source security (LinkedIn).
- Secure by Design principles
From CISA and international partners (CISA).
- MOVEit
Cl0p exploits a SQL injection flaw, 2,773 organizations by Emsisoft's count (CISA).
- Feross Aboukhadijeh on this show
Socket, later another angel (episode 36).
- Eliminate SQL injection
A Secure by Design alert from CISA and the FBI (CISA and FBI).
- The pledge
68 companies at RSA Conference (The Register).
- Corridor
Jack leaves CISA with more than 250 signers, and starts the company with Ashwin Ramaswami (CyberScoop).
- A seed round
Led by Conviction, after angels in March.
- 45 percent
AI-generated code introduced a flaw in 45 percent of tasks (Veracode).
- $5.4 million, and Alex Stamos
The seed announced, with Stamos as chief security officer (Axios).
- OWASP Top 10, 2025 edition
Broken access control first, injection fifth (OWASP).
- Episode 84 airs
This conversation. "Fully preventable for decades."
- A $25 million Series A
Led by Felicis, at a $200 million valuation (Forbes).
- Testimony
To the House Homeland Security subcommittee on AI security (House Homeland Security Committee).
- Coding agents, tricked
Malware through a support ticket (X).
- Argos
A model that picks codebase-specific guidance for coding agents (LinkedIn).
- Corridor joins Cognition
To work on Devin Security (Cognition).
What to do this week
Five things, in the order the episode suggests.
- Pick one bug class and close it at the source. SQL injection or authorization are the usual first choices. Make the safe way the default way in your framework, then add a check that fails the build when code goes around it.
- Teach your coding agents the same rules. Give them codebase-specific security guidance before they plan, not a generic rules file, and keep it current as the code changes.
- Put a security review on every pull request. Automate it, so it runs on every change without slowing anyone down, and save your people for the changes it flags.
- Validate before you prioritize. Find out which scanner findings are real, rank the real ones by exploitability, and feed the false positives back so the noise shrinks.
- Publish a vulnerability disclosure policy. It’s one of the pledge’s seven goals, it’s how Jack’s first bug got reported instead of exploited, and it takes an afternoon.