Matt (@wraakh) header
Matt (@wraakh) profile picture

Skeptic and science geek working with IT security.

Stockholm, Sweden
Joined June 2010
Photo by Matt (@wraakh)Photo by Matt (@wraakh)Photo by Matt (@wraakh)Photo by Matt (@wraakh)Photo by Matt (@wraakh)Photo by Matt (@wraakh)Photo by Matt (@wraakh)Photo by Matt (@wraakh)Photo by Matt (@wraakh)Photo by Matt (@wraakh)
I did some private research last week and let Claude write a bunch of code. A security scanner kept finding the same problems over and over. Nothing exotic — mostly habits. If something was missing — an owner, a check result — the code usually let the action through. It only said no when both sides were known and different. Tests with complete data passed, but real data is often incomplete. Guards that got added usually only covered the path we were looking at. Neighbouring paths to the same place were left alone. Authorization often stopped at “is this a valid caller?” without checking that the object actually belonged to that caller. You could get different errors depending on whether the object existed or not. Internal endpoints also ended up on public ports, protected only by a source-address check that a proxy in front made meaningless. Public endpoints that looked harmless could do expensive work — scanning a database or hashing a password — with no limit. The first fix was often a cache an attacker could bypass by varying the input. Claude also picked old, vulnerable dependencies without checking them. Pins, hash work factors and protocol choices reflected whatever was current in the training data. A version that looked specific could already have dozens of known advisories. The right technique landed in the wrong place. HTML-escaping where a URL scheme check was needed. A secret “protected” by a token that still crossed plain HTTP. Missing hardening headers that nobody noticed because nothing broke without them. Small inconsistencies piled up at the edges: wrong exit codes, raw subprocess output passed back, documentation describing behaviour the code no longer had. The common thread was that the code was written and tested for the scenario in view, by the same person who decided what “working” meant. Make sure you’ve covered these in your definition of done: - Missing inputs fail closed (not open) - Every path that reaches a protected state has the same guard - Ownership is checked, not just caller identity; request source is verified in the real deployment - Anonymous callers cannot trigger expensive work (or the response is cheap) - Dependencies, pins, work factors and protocol choices are current and checked, not remembered - Each check or encoding actually stops the thing it is meant to stop in that exact place An independent second pass catches most of it. Automated code review has to be part of the process from the start. Claude doesn’t make the same mistakes today as it did six months ago, and in six months it will make different ones. Other models make different mistakes. Use a mix of models and agents to review continuously, and make sure the agent actually looks in every relevant file. There’s not enough research on this front and we need to talk about it more. I’ll be discussing it at the SecureAI conference next week. See you there.
2
1
69
Early onset #vulnmageddon. vulnmageddon /ˈvʌln.məˌdʒɛd.ən/ n. A condition in which the established systems for identifying, triaging, prioritizing, and remediating security vulnerabilities become so thoroughly overwhelmed that they cease to function. It defines the era or crisis period where no existing process, tooling, or organizational model continues to work, and every practice for handling vulnerabilities must be discarded and reinvented from first principles.
1
1
22
We had an everything-is-broken day yesterday… again. PHP2Shell – two critical unauthenticated RCEs right in the PHP core – had been sitting there for over ten years. It wasn’t the project itself that found it. GitLab shipped another CVSS 10 unauthenticated RCE that had been around since 2014. Node.js fixed, among other things, a Brotli bug that can lead to RCE. eMudhra’s CA keys were so weak some of them were single-digit numbers. And today: a Signal bug that keeps the mic listening after you hang up. Many projects are already running Mythos, Daybreak and similar. But it’s not enough. The flaws are needles in a giant (hay)stack made up of many languages, more tools and runtimes, and even more dependencies that all come from different package managers. Open agents give you volume, not coverage. The ones who actually found these had tighter loops and more shots at the same code. We can’t patch our way out of this. We’re living in #vulnmageddon. Next comes #breachmageddon. People need to own their dependencies and actively minimize them. Remove one dependency per day if you can. Have an LLM reimplement only what you actually need and turn it into in-house code. One day you’ll have eliminated the supply chain risk. Maybe today is a good day to stop using PHP. (and come to our SecureAI.se conference 12:th of November if you are in Sweden)
1
1
174
A few seats left on the training courses if you happen to be in the #Stockholm area in November. Also come to the conference day the 12th! SecureAI.se
13
In the last Builders & Breakers podcast episode @hkashfi left us with a wish at the end of the last episode: better local models. For episode two, KUGG and I sat down with Thomas Olofsson, aka Skjortan, who has spent the past few years figuring out what those models can actually do in offensive security. Thomas brings a small box of local AI models to client engagements. The data stays on site. His agents help map systems, investigate findings, verify vulnerabilities and produce tests the client can run again later. We got into the practical details: hardware, harnesses, specialist agents, and why smaller models need the right skills and a clear threat model to do useful work. The part that stayed with me was where the conversation went next. If we’re getting much faster at finding vulnerabilities, we need to get much faster at fixing them too. A proper Builders & Breakers conversation with someone who is building this stuff and using it in the field. Builders & Breakers | Local AI for Offensive Security: Models, Harnesses and Human Judgment with Thomas Olofsson youtube.com/watch?v=xn5mIBaR…
2
3
30
4,330
Speaker announcements for this years SecureAI.se conference on November 12th at Nalen in Stockholm are under way. This year we are also hosting some relevant AI trainings the days before. Grab your tickets today while they are cheaper. See you there!
27
Matt retweeted
Tomorrow is the day. SEC-T's Community Day kicks off — see you all there. Münchenbryggeriet, Stockholm #SECT #infosec #seeyoutomorrow
2
3
420
Matt retweeted
SEC-T needs your papers! The 2026 CFP Round 1 closes 24 May, don't let that research sit in your notes! #SECT #CFP #infosec event.sec-t.org/sec-t-2026/c…
1
1
327
Matt retweeted
SEC-T Spring Pub CFP is still open. The prize? The General acclaim from your peers and a free ticket to SEC-T 2026! All we ask: show us something real. 15 minutes, in person, Stockholm on 20 May. Deadline: 13 May. #SECT #infosec sec-t.org/
1
1
403
Matt retweeted
Submitting to SEC-T 2026? Round 1 deadline is 24 May. Submit early and your talk gets another look in Round 2. #SECT #CFP #infosec event.sec-t.org/sec-t-2026/c…
2
2
941
NEW - EU reportedly plans to pass the controversial "Chat Control" legislation "without discussion" behind closed doors. disclose.tv/id/9ibxynfrek/
383
1,969
496
5,226
746,797
We must go deeper.
Well… this is not wrong…
311
3,517
478
31,779
6,034,695