CVE-2026-73615

CVE-2026-73615 is a quote-stripping bypass in your sandbox policy enforcement, but calling it a bypass obscures what it actually is: a structural mismatch between how your policy checker interprets a command string and how your executor builds argv from it. The policy evaluator treats quotes as literal characters to be matched against a blocklist. The shell executor treats quotes as syntactic directives to be stripped before argv construction. These are incompatible semantic models operating on the same user input — and they were likely never documented as a contract, which is why they diverged silently. Start by examining whether your 5.15.1 patch unifies the command representation or merely adds more quoted strings to the blocklist. If the patch only expands the blocklist, you're not fixed — you're patched. The vulnerability class remains open to future syntactic variations. What you want to see is one of two things: either the policy checker now tokenizes input exactly like the executor, or the executor now preserves quote structure for policy evaluation before constructing argv. Without one of these, the semantic gap persists. Beyond this specific CVE, audit your environment for the broader pattern. String-based policy evaluation paired with argv-based execution is a template that gets copy-pasted across systems — every sandbox, approval gate, or CLI tool that separates "check" from "run" likely carries this same latent gap. The recurrence risk isn't just about this version; it's about this architecture being reinvented in proprietary forks, internal tools, or new services that share no CVE lineage with this one. If you can't enumerate everywhere this pattern exists in your infrastructure, assume it's already there. The fix isn't better blocklists — it's contract tests that validate policy decisions against actual execution outcomes, every commit. Your policy checker and executor must agree on what "the command" means, or they must be architecturally isolated so their different interpretations don't matter.

Reviewed through automated stages and approved by a human before publication.