Appsec roundup - September/October 2026
New threat model tools, a PHP bug, chain-of-thought isn’t, and ... we’re not playing games here, or are we?
There’s been a lot going on at Shostack + Associates, so we’re combining the September and October roundups. This month leads off with two articles from an online symposium on the LawBlog Balkanization. Each presents an important insight into AI. The first is Professional AI’s Dual Trust Problem, by Claudia E. Haupt and the second is Ignacio Cofone’s Public Institutions Can’t Outsource Their Reasoning, about how a mix of over-reliance and non-explainability led to 26,000 families being accused of tax fraud, and leading to the resignation of an entire Dutch government.
The second is a horrific example of what I called
everything that is supposed to make a system like this accountable was in place. There was an approved algorithm and a procurement process behind it. Civil servants reviewed every flagged case. Procedures, escalation paths, and appeal rights existed on paper. None of those gave the institution the ability to answer for the outcome.
Trust is a system property, hard-won and easily lost. With that, on to the ways to preserve it!
Threat Modeling
- In Security Review at the Speed of Agents Vikram Ashtaputre talks about a system for review he built at RobinHood. Some interesting lessons.
- In
TALA is open-source, Alexander Wang releases TALA
(Terrastruct's AutoLayout Algorithm):
TALA is a novel autolayout algorithm designed with software architecture diagrams in mind. This means it's primarily an orthogonal layout engine, which more closely matches what you might find on whiteboards, rather than the DAG-based ones that grow in one direction.
TALA example #2 is a secure-enclave SoC, and note that the colors are not explicitly boundaries, which fooled me for a second.
Appsec
- There was a well-publicized bug in PHP and in We can now find flaws faster than anyone can fix them, Mihail Kirov writes about how an OSTIF-funded audit found the issue two years ago. (I’m an advisor to OSTIF.) There’s bitter, angry controversy over ChainGuard’s choice to release the patch early, either violating an embargo or responding to reports of exploitation. I’m not going to wade into that fight. What I will say is the fight exposes a lack of institutions who can provide authoritative information to inform ending an embargo, or even offer a call that will be respected.
- Dan Luu’s No reason for software to be slow starts: “...performance work that used to require... a rare set of skills can be done by anyone who can type a few sentences, which means that you can do all sorts of optimizations that used to be too expensive to be worthwhile”
- In Dear Open Source Developers: It's OK not to embargo, Josh Bressers writes about Curl’s decision to treat low and medium severity security issues the same way Curl treats other bugs: without secrecy. He makes great points about the cost of that secrecy.
AI
- In Reasoning Models Don't Say What they Think, a team from Anthropic demonstrates that “chain of thought” is not an accurate reflection of model calculations. (It’s nearly impossible to describe this without falling into anthropomorphization.) This finding is mirrored in METR’s Brief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident (“Brief” at 31000 words!) about which I hope to have more to say.
- In Agent Security is a Systems Problem, a long list of authors make a case for the titular claim, show that there's a set of tools that address agents, and discuss some of the challenges of doing so.
Cryptography
- In There’s a new way to break RSA that’s faster than anything we’ve seen before, Dan Goodin reports on new work by Nadia Heninger* and colleagues. Reduces attack complexity to 265 for RSA 1024, and 290 for 2024 against textbook RSA, but not PKCS or PSS. (Goodin describes Heninger as lead author; the preprint puts her in the middle.)
Games Received
Christmas came early with a box of goodies from Cybersec Games.
- Context & Cringe
- OWASP Cornucopia Web App Companion Set - 25ᵗʰ Anniversary Edition
- OWASP Cumulus for Cloud architectures
- Elevation of Autonomy for Threat Modelling AI Systems
- Insider Threat Kill Chain
- Security Architects
- Elevation of Privilege threat modeling workshop kit including EoP, three instruction sets (the method, an architecture and a facilitator’s guide), threat cards, a log, and a scorekeeping booklet.
Shostack + Associates News
We have lots of OWASP and ThreatModCon Plans, including a few seats remaining in our three-day Intensive Using AI covered in an extended blog post.
Image by midjourney, repeated from Oct, 2024