Articles by Niels
Activity
6K followers
-
Niels Provos shared thisThis week I tried to do something different with Claude. I stopped using it on any of my projects where the word security might appear in the repo and significantly reduced my overall usage - favoring GPT-5.6 Sol and GLM 5.3. As a result of my self censorship, I had a perfect week with Claude. Not a single false positive refusal. Did I learn that my experience with Claude is best when I don't use it as much? I am somewhat confused about the takeaway here.
-
Niels Provos shared thisAre you part of the growing club of cybersecurity defenders who have been locked out by the frontier labs from doing defensive cybersecurity research? I have a whole gallery of screenshots of false positive security refusals; often paired with a warning that I may be in violation of acceptable usage policy. Ironically, the labs themselves don't operate under these restrictions. Both Anthropic and OpenAI recently compromised companies because they themselves don't operate under the standards that are imposed on security researchers such as myself. Continuing the path of being in charge of my own destiny, I stood up a small DGX Spark cluster and with the help of Deepseek V4 Flash 0731 was able to conduct vulnerability discovery on libssh without being subjected to false positive security refusals. Getting everything working was a journey on its own. If you want to read more, including my findings with libssh, read the article below. If you are also part of the club of security defenders who are prevented by a frontier lab from using their models for defensive security research, I hope this is helpful to you. Open-weight models to the rescue and they continue to become more capable. There is hope for us yet :-) Link to the article: https://lnkd.in/gYxsGsZj cc: dan farmer
-
Niels Provos shared thisMy security education game now has a name: Exploit Margin. Thank you to everyone who voted in the poll. It got 79 votes and close to 40% were in favor of Exploit Margin. Exploit Margin is a business simulation about growing a technology company while persistent adversaries try to break in. You decide when to quickly grow the business, when to invest in security, and when to build infrastructure that makes entire attack paths impossible even when vulnerabilities remain. If you’d like to follow development or be considered for a future playtest, please join the list: https://lnkd.in/gwtPQ9eP Thanks to everyone who participated including the alternative suggestions. Some of them were quite funny ;-)
-
Niels Provos reacted on thisNiels Provos reacted on thisNiels Provos analyzed data breaches and found something that should fundamentally change how we think about security: 65% of breaches were preventable with just 3 security invariants, technical controls designed to categorically eliminate entire attack surfaces. The fundamental takeaway for security leaders is that we should stop trying to manage endless lists of vulnerabilities and buy dozens of point solutions. But instead, we should be asking: which entire classes of attacks can we categorically eliminate? To help you think through this yourself, Niels has been building the CISO Challenge, a gamified security strategy challenge to give you a new, more visceral way to experience the choices and consequences that define modern security. Link in comments. What are your thoughts? 👇 In case you missed it, catch up on the full conversation with Niels Provos on YouTube and podcast platforms. It was truly fascinating. You'll be blown away. Link in comments.
-
Niels Provos shared thisMy Claude weather continues to be stormy. Speed trended 22% higher today but my week still has 21 model refuses and 18 API failures. Try it yourself with: npx tokenenvy The tool continuously monitors your Claude sessions and keeps track of performance, refusals and outages for a longer time horizon than the native Claude log retention. Measured locally. Prompts stay private. #TokenEnvy.
-
Niels Provos shared thisEver wonder whether Claude Code is particularly slow today? Whether it's just you or everyone else as well? Token Envy is the answer. Analyze your sessions and share your stats. npx tokenenvy Claude Code took the scenic route that day: 38 effective output tok/s. How does your Claude Code speed compare? Measured locally. Prompts stay private. #TokenEnvy.
-
Niels Provos reacted on thisNiels Provos reacted on thisRemember the 27-year-old OpenBSD bug that Mythos found? Today, we'll hear from the guy who originally committed the code back in 1998, Niels Provos. He rebuilt the entire AI-driven discovery workflow himself. But here's the twist... he didn't need Mythos. Today we will see exactly how he did this and we will unpack what that means for you and the future of vulnerability economics. Niels also made the case for building security invariants, controls that eliminate entire attack classes without requiring human decisions every time. In his analysis of 70 real breaches, 3 controls would have stopped about 65%. If you're new here, hi and welcome to Security Mondays. Every Monday I bring you the best global security leaders with practical AI security guidance you can turn into action this week. If you have ever found value in this show, please do me a huge favor, engage with this post and follow along. This tiny action costs you nothing, but it has a huge impact the entire security community, because it helps bring the voices of some of the most incredible leaders in our industry to more people. Niels Provos is the founder of Security Blueprints and creator of Iron Curtain. He previously spent 15 years at Google as a Distinguished Engineer, served as Head of Security at Stripe, then Head of Security Efficacy at Lacework, and has spent decades working on vulnerability discovery and defensive security. If you've used a Linux server, a password database, or a honeypot in the last 20 years, you've probably encountered his work. He co-created bcrypt, the password hashing function, wrote libevent, built honeyd and the Systrace access control system and the StegDetect steganography tools. LinkedIn caps me at 15 minutes, so the full episode is on YouTube and wherever you get your podcasts. Links are in the comments. ♻️ If this is useful, please share it with your security, engineering, and leadership teams. Collective defense only works when we share what we know. Want to hear from more voices in AI security? Security Mondays is a proud partner and supporter of the AI Cyber Magazine with Editor in Chief, Confidence Staveley. Get your summer copy if you haven't yet and look for Security Mondays on the back cover. 🫶 If AI can find vulnerabilities faster than ever before, should we be spending less time finding them and more time making them harder to exploit? 👇
-
Niels Provos shared thisThis week's autonomous-agent incident produced a familiar recommendation: detect faster. Defenders will not win that race. An agent can run hundreds of attempts in parallel and never tire. The same workflow can target one company or a thousand. This risk has been materializing over the past year, yet most companies still make security trade-offs calibrated for human attackers. Instead, change what a compromise can reach. I reviewed roughly 70 public breach reports against three properties: mandatory hardware second factors, positive execution control, and no general internet egress from production. About 65% of those incidents could have been prevented or materially contained if those three invariants had been in place. A security invariant is a property that holds across every system in scope, with no local exemption and no dependence on someone noticing an attack. Most companies have fine-grained security controls. Far fewer enforce invariants. A compromised production service with no route to the open internet cannot establish arbitrary command-and-control or exfiltrate data directly. The agent can try that door ten thousand times. It stays shut. This is not another control to add to a risk register. It is a property you build into the platform. Each invariant eliminates a class of attack surface across every system at once, and the compromises that still happen stop at a boundary instead of spreading. Assume the attacker wins the first exploit. The presence or lack of invariants determines what that gets them. Read more: https://lnkd.in/gszVV3SF
Niels Provos replied to a comment
13h
Niklas Hallqvist good suggestion but it doesn't work: we detected that your musicality may be violative cyber use and against our acceptable use policy :-)
Niels Provos replied to a comment
13h
GPT 5.6 Sol is very good but has some idiosyncracies where in designs it wants to pin everything down to verifiable hashes - at least, I have encountered that multiple times - and it needs to be brought back to something more sensible.
GLM-5.3 overall seems very solid to me but the Z.ai coding plan has odd quota accounting. It seems to run out much faster than I would expect.
Niels Provos replied to a comment
2d
Nick Espinosa thank you, nick.
-
Niels Provos liked thisNiels Provos liked thisYet again, Niels is paving the way for us followers :-) Actually I am doing a lot of non-security related work, and still get demoted from using flagship models due to false positives in US AI model guardrails. China to the rescue, who would have thought that they would carry the flag of the free world?
-
Niels Provos liked thisNiels Provos liked thisThirty years ago, vulnerability research meant one person reading the right code. Today, AI can do it at scale. Niels Provos shares how vulnerability discovery is no longer a frontier-model problem; it's an orchestration problem. https://lnkd.in/g_jnGgfB
-
Niels Provos liked thisNiels Provos liked thisThe breaches will continue until incentives align 🤕 I just stumbled across a letter signed by OpenAI, Anthropic, Amazon Web Services (AWS), Google, Microsoft and 100+ companies calling for a "global surge in cyber defense as AI capabilities advance worldwide". It opens by stating we have a limited window to strengthen cyber defenses, and in the coming months AI-enabled cyber attacks driven by increasingly capable models will impact society, including critical infrastructure and institutions. Let me say, I agree with them completely, this is coming for every aspect of our digitally enabled society. But, I also want to point out a few things: Many of the signatories are cloud providers, frontier labs, and cybersecurity vendors with deep rooted incentives in pushing to make cyber a business leadership priority because it comes with the expectation of increased funding/spending and as a result, cyber revenues. We can't "will" our way into being more secure, as much as I wish we could. This is why efforts such as Secure-by-Design under Cybersecurity and Infrastructure Security Agency led by amazing folks such as Jen Easterly, Bob Lord, Jack Cable and others fell short. The prior National Cyber Strategy (NCS) started down the path of software liability and putting the onus on software producers, but as James Dempsey has stated, Software Liability is Cyber's 3rd rail - and it quickly fizzled out. Cyber is rightfully considered a market failure by many, and competing incentives and priorities such as speed to market and revenue will almost always be placed ahead of cybersecurity. Until market forces (impacting business revenues, valuations and social stability), or regulatory changes take place, this simply won't drastically change. Do I *wish* we took security more seriously? Of course. But, that's not how any of this works. https://lnkd.in/edrvJcnt
Experience & Education
Publications
-
Framework for Detection and Measurement of Phishing Attacks
Workshop on Rapid Malcode
Recommendations received
-
LinkedIn User
7 people have recommended Niels
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