Arc Sentry catches Crescendo, LLM Guard misses

The interesting part is the layer, not the score. If the claim holds up, session-aware whitebox monitoring is materially different from text classifiers for attacks that are designed to look benign turn by turn.

• –LLM Guard is facing the wrong problem shape here: Crescendo is built to evade per-turn text checks, so independent prompt scoring is structurally disadvantaged.
• –Arc Sentry’s residual-stream approach matches the failure mode better because the attack is about gradual state drift, not explicit toxic wording.
• –The headline benchmark is still vendor-run and narrow, so I’d want independent replication, calibration details, and real-world false-positive data before treating the 92% claim as settled.
• –The Arc Gate reference matters because it suggests the same stability idea is being extended from open-weight, whitebox monitoring to hosted API governance.