VLA detects covert AI-agent coordination

VLA tackles a real blind spot in transcript-based oversight, but its results remain controlled evidence rather than proof that spontaneous covert protocols can be detected in production.

• –Links private latent handoffs to public actions through shared event identifiers, enabling causal rather than purely correlational monitoring
• –The neutral-only training objective is promising because it avoids relying on a catalog of known attack codes
• –Performance falls to 0.854 AUROC for heterogeneous model pairs, highlighting the difficulty of aligning incompatible latent spaces
• –The auction benchmark uses receiver-primed, fixed latent attacks, so it does not demonstrate spontaneous emergence of covert communication
• –Whitebox matched-neutral steering is powerful but requires deep runtime access that many multi-agent platforms will not have