CVE-2026-73614 in ClaudeHookBridge presents a textbook case of evaluation pipeline divergence: the denyPatterns security check truncates input at 500 bytes, while the command executor processes the full payload. An attacker who places malicious content beyond byte 500 can bypass denylist enforcement entirely — the filter sees a clean truncated view, the execution engine runs something else entirely. The 8.8 CVSS reflects genuine impact. The 0.00358 EPSS tells you this hasn't weaponized yet. That gap is the signal you should be reading, not the noise.
This isn't a sloppy input validation bug — it's architectural. The security layer and the execution layer operate on different representations of the same data, and that divergence was almost certainly introduced by a developer reasoning about reasonable input lengths, not by anyone asking 'what happens past this boundary?' That pattern — evaluation divergence — is endemic to tooling and CI/CD pipelines where security gets bolted onto execution pipelines after the fact. The 5.15.1 patch will likely extend the truncation boundary, and that will probably create a new bypass window at a higher offset. The class of vulnerability doesn't die with one fix.
For detection engineering, the problem is acute. Your SIEM likely ingests whatever the denyPatterns evaluation saw — the truncated view. The malicious content past byte 500 is invisible to standard logging. To catch this, you need to instrument the execution layer itself: capture raw process arguments at runtime, not the filtered view. Specifically, audit your logs for commands exceeding 500 bytes that executed when your denylist should have blocked them. That's the forensic signature of this bypass — and it's one your current alerting may be blind to.
Two things to watch. First, the EPSS-CVSS gap will likely close once someone releases a generalized finder that automates payload construction across similar pipelines — this pattern has done exactly that in HTML sanitization and HIDS evasion. Second, your version inventory matters: organizations on pre-5.15.1 builds have a silently failing security control and no CVE notification infrastructure reaching the operators who configured those denylists. Check what version you're running, because the vulnerable code path left no failed-deny trace — the logs looked clean even when the bypass worked.