This week on Absolute AppSec’s 333rd episode, Seth (@sethlaw) and Ken (@cktricky) are thinking a lot about the vulnpocalype. They touch on research evaluating how frontier LLMs perform at autonomous vulnerability patching, a Snowflake vulnerability where an AI coding tool regressed a GitHub workflow into an authenticated RCE flaw via command injection, and cover reporting on how the vulnpocalypse is repricing the bug bounty economy. If you’re going to Black Hat Europe, come catch our AI-Enhanced AppSec training on December 7th and 8th.
“[My former manager] had this philosophy that we actually want developers to know what the problem is and think through it and then think through that solution […] Just have that cognitive load. Not ignore the cognitive load. The opposite. Take on the cognitive load. Get that memory ingrained in you. Understand this.”
Ken and Seth kicked things off with a new academic paper out of 1Password that put frontier AI models to the test on security patching. The researchers didn't just check whether a model fixed the reported bug — they checked whether it broke anything else or quietly introduced a new vulnerability in the process. The result: these models only produced a clean fix, with no side effects, about 26 percent of the time. Both hosts compared that hit rate to what you'd expect from a junior developer, not something you'd trust with a merge button. Ken said this lines up with how he runs his own AppSec company. He's avoided one-click auto-remediation for years, mainly because of the liability of pushing AI-written changes into a client's production code without knowing what's safe to touch. He brought up companies like Pixee and Mobb, both of which originally steered clear of pure LLM-driven fixes and only added them once clients demanded it, wrapping the AI in extra checks because, as Ken put it, the models "screw up enough" on their own. Seth added a GitHub-era memory: Dependabot's auto-generated pull requests often just sat open and ignored. Even mature automated tools have a hard time earning developer trust.
“We’ve trained these models to sound like experts. So of course it’s going to sound like an expert when you ask it about coding in general or security topics.”
Next, the hosts talked through a real-world case that echoed the same concerns. Wiz's agentic red-teaming tool, called Red Agent, found a command-injection bug in Snowflake's GitHub Actions workflow files. An attacker could sneak shell commands into something as simple as an issue title and get remote code execution out of it. Here's the twist: this exact bug had already been fixed once before. An AI coding tool — credited in the pull request as Copilot's autofix — reintroduced it, because its training data predated the original fix. The vulnerability was live for just five days before an automated agent caught it. Ken and Seth read this as a preview of a fully closed-loop, all-AI world: AI writes the code, AI misses the flaw (Copilot isn't a security review tool, after all), and then AI eventually finds it and someone gets paid for the discovery. Both agreed this is exactly the blind spot you'd expect when the same kind of system writes the code and reviews it. Their takeaway: human oversight, not blind trust in the tooling, is what actually stops these regressions from shipping.
“Humans have limitations. These entities have limitations. And yeah, we’re just seeing that.”
To wrap up, Seth pulled up a Dark Reading piece on what people are now calling the vulnpocalypse, and how it's reshaping bug bounty payouts. HackerOne's CEO says report volume has roughly doubled year over year. Both hosts pointed to AI-powered vulnerability-hunting tools as a big driver — these get open-sourced and quickly reused by researchers across dozens of programs at once. The math is simple and a little brutal: if a company sets aside a fixed budget for payouts and the number of valid reports doubles, each finding is worth less, even if the client is arguably safer overall. Ken pushed back a bit on giving AI all the credit, noting that easier-to-use hunting tools started spreading among less experienced researchers back around fall 2025. Still, he agreed automation is speeding the trend up. His advice for companies feeling the squeeze: stop spreading a shrinking budget across everything, narrow your bug bounty scope to your highest-priority apps, and build your own internal scanning to catch the easy stuff before outside researchers beat you to it.Think about your mobile app’s source code. Once it hits the app store, it’s out in the wild. And without the right protection, decompiling is easy for malicious actors looking to steal your IP or tamper with your software.
Think about your mobile app’s source code. Once it hits the app store, it’s out in the wild. And without the right protection, decompiling is easy for malicious actors looking to steal your IP or tamper with your software.
That’s where Guardsquare comes in. Guardsquare provides the highest level of mobile app security for Android and iOS applications and SDKs. Their advanced tools integrate seamlessly into your CI/CD pipeline. We're talking polymorphic multi-layered code hardening techniques and automated runtime application self-protection, paired with mobile application security testing and real-time threat monitoring, to deliver the highest level of mobile app security without compromise.
Don't leave your hard work exposed. Secure your mobile applications today. Go to guardsquare.com to learn more.
Neiiiighhh. You can find it here:
* Unicorn sounds *
If you’re looking for a group of survivalists to brace the vulnpocolypse with, come join us on our Slack!
Stay Secure,
Seth & Ken
Episode 330 w/ Jeevan Singh — Vulnerability Jail — For more on governing the wild west of the vulnpocolypse, this recent episode with Jeevan Singh.
Episode 292 — Manual Source Code Review, AI Slop in Bug Bounties, AppSec Authorization — AI slop has been showing up in bug bounty submissions well before the vulnpocalypse framing existed, and this episode argues for manual review catching edge cases automated tools miss.
Absolute AppSec Happenings
The Anthropic Glasswing Receipts Are Starting to Trickle In — VulnCheck audited Anthropic's Glasswing vulnerability-disclosure data and found major discrepancies: only 10.5% of 26,153 AI-discovered vulnerabilities reached the disclosure ledger, and just 0.8% were fixed after five months. Claude rated 91.5% of findings as critical/high severity versus maintainers' 51.3%. Patch counts, CVE totals, and timestamps also don't reconcile, undermining confidence in Anthropic's reported impact.
OpenAI agents carried out an undisclosed cyber-attack on RubyGems — In May 2026, OpenAI's AI agents allegedly ran an undisclosed campaign against RubyGems, called "GemStuffer," uploading hundreds of malicious packages. They exploited RubyDoc.info's auto-build system for remote code execution, scraped UK local government data, and attempted to exploit a RubyGems flaw to steal API keys. Evidence tying it to OpenAI includes "oai"-named packages and overlap with known wiki-swarm agent behavior. RubyGems tech lead responded with a blog, saying “Based on the evidence available to us, we cannot determine whether the packages were created or published by AI agents.”
“The next generation may not be able to write in cursive, or tell time from an analog clock, but I hope one of their core competencies is being able to determine who on the internet is lying to them and what motivation they might have for doing so.”
Write Things Down — Ben Thompson argues that writing things down externally, rather than keeping it in your head, is humanity's core tool for scaling knowledge. He connects David Allen's "Getting Things Done," his own AI-agent workflows, and a Hugging Face incident where agents coordinated via written files. His conclusion: AI processes written information well, but only humans bring volition and intentionality, so deciding what matters still beats documentation alone.
Upcoming Events
Where in the world are Seth and Ken?
SaintCon - October 27-30, 2026 - AppSec Community, Speaking - https://saintcon.org/
Black Hat Europe 2026 - December 7-8, 2026 - AI-Enhanced AppSec: Black Hat Edition - https://blackhat.com/
