On Episode 337 of Absolute AppSec, Seth and Ken (@sethlaw and @cktricky, respectively) are joined by Kurt and Cameron of the Coffee, Chaos, and ProSec podcast for their quarterly collab. To find all our other episodes with Kurt and Cameron, head over to https://www.youtube.com/@AbsoluteAppSec/streams, or find us wherever you get your podcasts. Please be sure to leave us 5 stars on Apple Podcasts or Spotify.

These days, everyone’s feeling tired. Kurt described the news cycle as the same story on repeat— another agent getting loose and biting someone— and admitted nobody, himself included, has gotten sandboxing right. If Anthropic is reportedly siloing its own models because guardrails are that hard, what chance does everyone else have? Ken read the reported talks around OpenAI acquiring Trail of Bits as frontier labs hunt for expert knowledge that public training data can’t supply, effectively commoditizing consultants’ work. The counterweight, Kurt and Cameron argued, is a move toward small, specialized, open-weight models running locally, driven by shrinking token budgets, privacy, and cost. Seth’s frustration was sharper: Redpoint routinely gets blocked by frontier-model safety filters (even while drafting this episode’s agenda) while attackers freely use unrestricted models. When vendors cut enterprise AI deals with the C-suite, Kurt’s advice was to insist on the cyber models too, because, as Seth’s two-day, false-positive AWS account suspension showed, every delay now hands attackers the advance. 

The biggest debate centered on detection versus remediation. Ken observed that moving from deterministic scanners to agentic tools usually triggers an explosion of newly discovered bugs, which makes progress hard to measure. Cameron pushed back, arguing that teams already sit on a mountain of known risk, so the real win is remediation capacity. If a team finds four thousand issues a month but now fixes eight thousand, it can finally start hunting the interesting bugs. Seth noted that most mid-sized clients aren’t there yet. They still judge AI tools against legacy DAST scanners that report a hundred findings, ninety-nine of them false positives. Cameron challenged practitioners to ask what their value really is, because a CEO who learns that twenty-nine thousand of thirty thousand findings don’t matter won’t be impressed. The bright spot is threat modeling. Classic frameworks like Microsoft’s STRIDE and DREAD used to mean thirty-page documents nobody read, but Kurt called AI-assisted threat modeling “light years better.” Seth said Redpoint has run more threat-modeling engagements this past year than in the previous six or seven. They agreed that defenders’ edge is context. They can give agents white-box knowledge and earlier pentest reports, which attackers simply don’t have. 

Ken closed with a hot take from his essay on the Next Generation of AppSec: Is the security industry built to solve problems or to sell solutions? Many practitioners, he said, have tied their identity to finding and exploiting bugs, yet AI absorbing that work frees them for everything they never had time to do. Kurt called it a bit of both. Cameron dismissed compliance questionnaires that ask whether you’re “Mythos-ready” as security theater. Mythos is Anthropic’s security-capable model, which rattled security stocks this year. Ken added that buyers share the blame, since many still believe the bigger number wins. On auto-remediation, Ken doubted that opening two thousand pull requests would get more than a tenth merged, because developers want control. Cameron’s answer is for vendors to ship token-efficient, human-in-the-loop remediation skills that developers can run, tune and learn from, noting that bloated skills simply get switched off. Seth shared that Redpoint now includes agent-ready prompts with its findings and is exploring skills that eliminate entire vulnerability classes. As Cameron put it, the shared goal is escaping the firefighting doom loop and systematically burning down risk. 

Here are some recent memes from our Slack! We’re fun, we promise.

Stay Secure,

Seth & Ken

Coffee, Chaos, and Absolute AppSec: The State of ProdSec | Q3:26 - Ep 61 - The other side of our quarterly collab. Here, we chat all things vuln counts as value metric, bug janitor careers, frontier lab lock-in, and whether AI is enabling security teams or burying them.

PHP2Shell Step By Step: From Engine Internals in C to Real Exploit Code — Ariane Djad’s deep dive covers PHP2Shell (CVE-2026-109618 and CVE-2026-109619), a remote code execution chain in PHP 8.4.26. When an upload pushes PHP past its memory limit, cleanup gets skipped and stale pointers are left behind. The author uses those pointers to leak heap and code addresses through session cookies, then plants a fake structure that runs system(). The offsets only work for one specific build, and the article doesn’t cover patches or fixes.

Leading AI models are biased against certain accents when scoring interviews, presentations and tests — University of Washington researchers found that AI audio models from Google, OpenAI and Qwen rate English speakers differently depending on their accent. The models scored identical scripts lower when spoken with Chinese or Nigerian accents, and highest with American or British accents. This happened even though the models transcribed the speech accurately. A follow-up test with 52 real speakers showed the same pattern. The authors warn this matters as AI increasingly evaluates job interviews and English proficiency tests.

OpenAI CEO says it’s ‘quite terrible’ for your PC to have Ram — Sam Altman says data centers offer better utilization — In a Cheeky Pint interview, OpenAI CEO Sam Altman said it’s “quite terrible” for people to have “a bunch of the world’s limited supply of DRAM” sitting in their own computers, used only “5% of the time.” He was then heavily clowned on in our Slack.

Where in the world are Seth and Ken?

Black Hat Europe 2026 - December 7-8, 2026 - AI-Enhanced AppSec: Black Hat Edition - https://blackhat.com/