The beast needs a cage: What's next for App Sec post-Mythos
Summary: PortSwigger's Dafydd Stuttard writes that the traditional AppSec value chain of finding, proving, and fixing vulnerabilities will be automatable end-to-end, and that open-weight models will likely match today's frontier capabilities within a few months. He argues LLMs are non-deterministic and cannot self-govern, so agentic offensive tooling needs a deterministic safety and governance layer with human-in-the-loop decisions and audit trails, and he cites James Kettle's agentic implementation of his research methodology finding a novel attack technique and stealing an API key from a bank while he slept.
What's compelling: Stuttard's point is that as model capability gaps close, the structural limitations get more significant, not less - more capable models take bigger, more consequential actions with less oversight. An agent that compromised a bank overnight without its operator knowing the target was in scope is the concrete stake for anyone planning to hand offensive tooling to a model.
Read the original at PortSwigger Research ↗
- Topics
- Vulnerabilities & Patching
- Related
- Attackers Exploit UnrealIRCd Flaw That Enables Server-Level Command Injection Cyber Risks · Nov 14, 2026U.S. CISA adds Check Point and UnrealIRCd flaws to its Known Exploited Vulnerabilities catalog Cyber Risks · Nov 14, 2026PostgreSQL Patches 52 Flaws, Including 22-Year-Old Bug Enabling Authenticated RCE Cyber Risks · Nov 13, 2026