
Today is the day we finally get to announce that we closed funding of our seed round for our new company Evidence. This day represents over a decade of brainstorming and years of gathering data and talking to design partners.

Managing Director - Grossman Ventures. CTO of Root Evidence. Defender of others' privacy, Author of AI’s Best Friend, Advisor, often found joking.

Today is the day we finally get to announce that we closed funding of our seed round for our new company Evidence. This day represents over a decade of brainstorming and years of gathering data and talking to design partners.

Hi, yall. I’m a cowboy coder. I write code in prod, I press save, and I test right there alongside my users at the same time they do. I’ve written code this way for 30+ years. I do it because it saves me time, and I hate managing complex environments with lots of test/staging/dev servers. I don’t even like managing servers at all. And now, with AI, many of you are doing it too. So let me give you some tips to reduce the chances of completely screwing yourself over. 1) Make small and reversible changes. Ideally really just a few lines or a single function at a time. I like to make copies of functions, test the new copy and quickly revert back when I’m sure it’s working. Big changes are hard to debug. 2) Use feature flags, version flags, or something to make it easy to flip back to older versions. This is harder to do when you are working with databases, but still doable, if you want to run parallel databases for a little while until you’re sure it’s working okay. 3) Don’t push and call it a day. Be prepared to spend the rest of the day fixing whatever you just broke. So ideally, push in the morning on a quiet-ish day. 4) Treat the code as broken, even if you are 100% sure it’s fine. It’s not fine. It’s almost never fine. Even the simplest changes often cause the most havoc. Don’t treat it like it’ll be just fine. Treat it like it’s broken and you’ll be better off. 5) Comment your code. I know, I know… trust me. I can feel you yawning already. But sometimes comments are hilariously useful. I had one comment that was a page long explaining how the function I wrote is the most efficient way to do it and how I spent over a day trying to make it better only to find out the RFC was wrong and then got stuck doing it an incredibly stupid but highly efficient way. Later on, other devs found that code, were sure I was wrong, ignored my comment and proceeded to spend almost a month trying to figure out how to make my code better. All I had to do was read it again to remind myself why they’d never improve my code… code I had long forgotten about and why it was written that way. Needless to say, my code persists to this day. 6) Have some standards. There are some things that just shouldn’t be done in prod, like anything that touches money, or deletes large chunks of data programmatically, or really anything offensive information security related. You want to do those things much more carefully. But if you do want to go whole-hog, at least roll it out a few percent of your users at a time. 7) Logs! Just log it all. Removing logs later is easy. Debugging without logs is a nightmare. And test that the logs are logging what you think they should be. 8) If you’re not going to review the code that AI writes, at least review the use and abuse cases, because AI is notoriously bad at finishing those to any level of utility. Make sure the behavior isn’t junk and tell the AI you will be reviewing that code personally so it had better do a good job. 9) AI is like a very good but very junior engineer. They know the syntax, but they have no idea what the product needs to be. So, spend the time to write really good product requirements or user stories. AI will be able to follow that, and some future QA engineer can know if the AI did a good job or not, if you have good requirements. If you have AI write them for you, you owe it to yourself to review them though, because they’re probably a hot hallucinated mess. 10) Backups. Backups, backups, backups. And test the backups. And that’s it! Enjoy being a cowboy coder! For all the shit cowboy coding has gotten over the years, it’s about to see its renaissance, due to AI. God help us all! Yeehaw!

BREAKING: The US Federal Aviation Administration (FAA), the agency responsible for civil aviation safety and air traffic regulation in the United States, has issued a nationwide ground stop due to widespread radio frequency outages affecting multiple air traffic control centres. Departures are being held until further notice.

Incident response firms are turning away work. A lot of it may have no attacker behind it. One investigator told Robert “RSnake” Hansen that in a run of crypto cases, 24 of 47 real incidents came from an AI agent that went off the rails. The agent caused the outage or moved the money out itself. That would explain why breach reports keep climbing while insurers see no meaningful rise in losses. Most DFIR analysts want to solve crime, not clean up after AI slop. Read Robert’s full blog: na2.hubs.ly/H08--NN0

The Vulnpocalypse - Good News and Bad News - The bad news is we are inundated with new vulnerabilities due to the increased volumes discovered by AI tooling. - The good news is many organizations are starting to handle this with increased speed, prioritization, and existing methods of defense in depth. - The bad news is there are also many organizations not doing this, and worse, are sitting on a whole ton of software, configuration, and structural/policy vulnerabilities from prior years. - The good news is many of these targets have not been exploited due to attacker’s resource constraints. - The bad news is that this might not hold as attackers industrialize their capabilities to turn more targets of opportunity into actual victims. - The good news is, apart from scams, crypto/Web3, and some parts of the ransomware kill chain, it appears that such industrialization is not helping monetization (yet). - The bad news is attackers are an innovative bunch and not likely to give up and go set up a bakery, or go back to drug dealing. Besides, chaos may still ensue as agent swarms simply exploit vulnerabilities anyway. - The good news is defenders can use all this AI capability for defense. We are seeing the potential that many significant code bases will have dramatically lower vulnerabilities levels than ever before due to the “deep cleaning” (h/t @argvee) with the very same tools we are concerned attackers will use. - The bad news is getting a code base “clean” doesn’t mean the updated packaged software will arrive quickly in all the places where attacks will target. The good news is……..? philvenables.com/post/the-vu…

This whole line of agentic deception to waste the adversary time/tokens shows a lot of promise: agentic.tracebit.com/factory

If your code is downloadable (as in a mobile or desktop app) it can be reverse engineered. A great use case for LLM-enabled adversaries, and makes webapps interesting again, because attacking them can scale with secrets derived from the decompiled apps.
Called it. Reverse engineering of mobile apps is now much easier and can be used to target their supporting back-end APIs, possibly using API tokens in the app. The right design is an authenticity boundary, where reaching the APIs requires attestation of authentic, published Android or iOS app.

Open question. I have my theories, but I would love to hear other people's theories on this. When we look at the amount of exploits for which there is a documented known exploited vulnerability (KEV) listed, compared to the number of exploits that have exploit code, it is clear that they are uncorrelated. So why is that? Why do hackers seem to completely ignore most vulnerabilities for which there is exploit code? The current stat for all combined public KEVs is only around 2.1% of code with exploit references is ever exploited and that number is declining as a percentage due to all of the new CVEs being produced in the last three years with the help of AI.

What AI is actually worth to an attacker? On October 14, @jeremiahg and @RSnake sit down with Michael Aret, a CISO in the defense sector for a nation state, to discuss how nation-state actors actually use AI, where it genuinely amplifies their capabilities, and where its unpredictability is a liability for an adversary whose operations fail the moment they're noticed. Webinar: AI and Nation States Wednesday, October 14 at 9am PT Register: na2.hubs.ly/H088Dwh0

We're now over 400k CVEs and counting. Update your talk tracks everyone! Onward and upward! cvedata.com/

Rare data within LLMs don't typically surface because LLMs produce the statistical mode of their training data. That works great if everyone is right. However, if they're not right, it's terrible.

Proud SEO, never forget your roots! #fellowship #strong #g50

Every AI hype dealer in cybersecurity should be forced to watch this. @jeremiahg is bringing the goods and the data. @rootevidence
AI is creating a lot of things that aren't real, including the supposed spike in zero days.

The Vulnpocalypse was supposed to be here by now... so where is it? In our new research, we traced every confirmed exploitation of a published CVE since 2018. On Wednesday, Sept. 2, @jeremiahg and @RSnake are hosting a webinar to walk through what the data actually shows. At the end of the hour, you'll be able to answer: → Whether AI is really growing your backlog faster, and what that changes about what you patch → How to tell the 1.5% of CVEs that get exploited from the ones that never do → Whether you should be planning for more zero-days than you were three years ago → Why the median patch-to-exploit window is the wrong number to build an SLA on → Which of your CVEs get attacked inside 30 days, and which sit untouched for over a year → Which vendors and vulnerability classes are worth most of your patching hours Attendees will get access to the full data in our upcoming Vulnpocalypse Report before it is publicly released. 🎙️ Webinar: Rumors of the Vulnpocalypse Have Been Greatly Exaggerated 📅 Wednesday, Sept. 2 at 9am PT 🔗 na2.hubs.ly/H07p4dw0

Something just came to my mind that @jeremiahg and @RSnake have done a lot of case deep-diving with css tricks back in from ha.ckers community days
I stole an Outlook password with nothing but CSS inside an email 👀 Whitepaper below 👇

Which vulnerabilities are actually going to cost us? The answer isn’t “all of them.” That’s one of the central ideas behind “The End of Guessing,” a new book from Robert Hansen @RSnake and Jeremiah Grossman @jeremiahg . Instead of asking security teams to chase every vulnerability, the book explores how evidence of real-world exploitation, adversary behavior, and business impact can help organizations focus on the vulnerabilities that will lead to financial loss. If you're attending Black Hat, stop by the Nucleus Security booth (#5533) to meet the authors and get your copy signed. 📍 Aug. 5 | 2:00-3:00 PM 📍 Aug. 6 | 10:00-11:00 AM #BlackHat #BHUSA #CyberSecurity #VulnerabilityManagement #ExposureManagement