General Application Security
More Burp Extensions
Burp Suite Extensibility Month ended back in May, giving rise to Burp is fine the way it is month. Now it’s July, but if you are as behind as I am, then you are only now just reading The 2026 Burp Suite Extension Awards. People really are using Burp in all kinds of ways, but nothing mentioned is likely to be new to you if you have been using Burp extensively for years. Interestingly, some of my most used extensions are not included.
Android Lock Screen Rate Limiting Changes
Android 17 is bringing more robust rate limiting changes. You can read about them here. Citing a study on PIN guessability from 2020, the primary changes include far fewer total attempts (20) and a counter that ignores duplicate attempts. The timeouts are roughly tripled each attempt, which creates oddly long timeouts after a number of failed attempts. Do we really need a 9 year timeout before we can submit our final guess? It is not clear what purpose some of the final timeout periods serve.
Developers Are Still Letting LLMs Loose
Anthropic wants to know if you’re Chinese and I am sure this will not be the last from the AI industry to extract information from its customers, which is why people really need to stop doing this:
Developers like me give Claude Code full filesystem and significant shell access so it can do its job.
Exploiting SSRF
This is a fine write-up of the methodology for identifying internal services with an SSRF capability, though the LLM-y nature of the writing is a major distraction.
Sometimes it is the residue of partial decommissioning
…
That was the pivot. The value of the SSRF was no longer theoretical. It was now possible to identify internal systems that were invisible from the public internet and interact with them through the vulnerable host.
Come on. Perhaps I am wasting my time as it feels more and more like we are doomed to an ecosystem of increasingly obviously AI generated infosec articles.
What Is a Vulnerability?
In my experience, a surprisingly few number of people working in infosec are able to articulate what they mean by terms like vulnerability, weakness, risk, bug, and so on. I think this is mostly a result of the lack of formalized education in the field (or lack of quality education) and having to accept an operational use of the concepts without thinking about their broader meaning (for example, even if someone working in vulnerability management cannot define vulnerability, they still know where they come from and what to do with them). Realistically, conflicts do arise because practitioners do not understand or do not having a coherent working definition of these terms. For most practitioners, MITRE’s vulnerability theory is probably one of the better guides, but even this document is not fully consistent with other industry authorities.
There is a post circulating in all of the popular appsec newsletters and podcasts (therefore it must be in this one): The Vulnerability Identity Crisis. The authors propose a new working definition for the term vulnerability. Their focus is primarily on the failure of CVE and CVSS systems to fully capture the properties of a vulnerability. Substantial details are missing compared to their full presentation on the subject. Perhaps the full talk will be more informative when released.
How LLMs Identify Vulnerabilities
An interesting experiment examined the value of repeatable patterns in the efficacy of LLMs identifying vulnerabilities, but I would hope you would expect this behavior; LLMs are pattern machines. What I find more interesting and – to some degree – distressing is the pervasive misunderstanding of the relationship between models, their outputs, and truth. The author makes a number of statements on the model’s propensity to lie or confess, but language models make no distinction.
The author concludes that “you cannot treat the model’s own explanation of its process as data about its process“ but this is partially incorrect. Perhaps part of the confusion comes from understanding the output as an “explanation” but the model does not distinguish explanatory content from other output; we apply our own meaning. The model’s output is the artefact of its process.
Crucially (and in an effort to help convey my point), the author’s conclusion is misguided: “Only in retrospect, Opus spots the similarity.” Rather, the conclusion should be “Only after additional prompting did the model output a reference to the known vulnerability.” We do not know the extent to which information relating to the vulnerability in questions influenced the model’s weights such that it arrived at a solution to the challenge even if the specific vulnerability is not referenced in the output.
For people who think of LLMs as magical thinking machines, it may be valuable to review analyses of how LLMs write. Or perhaps something more grounded in theory:
Even if we may have more in common with these models than we like, anyone who understands the first thing about how they work can see that attributing to them freedom, self-consciousness, or other powers of the humanist subject is a simple misunderstanding—all the more so if we do not possess such powers in any purity ourselves.
I Was Blocked On LinkedIn By A Trufflehogger
In case you can not identify it, an LLM is being used to synthesize information available about me in order to further a sales effort.
“Agentically Take Action”
Apple is introducing a new feature to mitigate risk from compromised passwords. I have not examined this in detail, but I am hesitant to let anything “agentically take action” on my behalf to manage account credentials. Meanwhile, Apple’s private email system has been leaking email addresses, which is the one thing it’s not supposed to do.
On the Microsoft side, the Windows team has finally decided to commit to quality. They’re doing this by once again integrating LLMs into the Windows ecosystem. The new system is based on OpenClaw, which has been doing great keeping users secure. I haven’t used OpenClaw or MS Scout, but based on planning documents, I have to assume it will function like a digital slot machine and provide everyone the same experience that developers have right now. That said, I am not so sure this will make a meaningful difference on an operating system that already feels non-deterministic without LLMs.
Leaks from Public Formatters
No one should be surprised of this, but here we are.
Neat Leak
Here is an interesting Chrome information leak.
LLM-Driven Reverse Engineering
There is no doubt that LLMs can accelerate many tasks in the reverse engineering process for vulnerability discovery. Here is a write-up where the authors basically just pointed Claude at a zipped binary. At the same time, there may now be the first legal case arising from inaccurate LLM output from a reverse engineering process.
HTTP QUERY
HTTP has a new method: QUERY. Summary here and RFC here. What is QUERY? “GET with a Body.”
WordPress Is Old and So Are The Systems Running It
Censys crunched some numbers on detectable WordPress instances, finding that over 70% ran outdated PHP versions, with 14% on the latest patch of WordPress. Unsurprising.
Malicious Copy Protection
The Opera browser has introduced a copy/paste protection mechanism that issues warnings when potentially malicious content is identified.
Standards
OWASP MASTG v2.0
Version 2.0 of the MASTG has been released. Release notes here. The MAS team does an incredible job working on their suite of standards, so this one is worth checking out if you work in MobSec.
OWASP AISVS
The AI-flavored ASVS for LLM-integrated applications (“AI-enabled”) has been released.
Connect
Respond to this email to reach me directly.
Connect with me on LinkedIn.
Follow my YouTube.
RSS feed here.
