New Technique Medium TD-2026-0129 For AI Eng, AppSec, CISO AI-assisted
Repeat-to-Reveal attack recovers token multisets from obfuscated KV caches (KV-Cloak)
The paper shows that shuffled obfuscation of KV caches loses security when inputs contain repeated tokens, because repeated tokens yield identical Value rows. The attacker can recover an equivalent denoiser from its own probes and reuse a one-time calibration across requests sharing a secret epoch. The authors report 100 percent token-multiset recovery on GPU for the official KV-Cloak defense and also confirm the issue on two other open-source defenses.
Why it matters
Infrastructure operators and co-tenants that share KV caches can recover token-level content from cache blocks, exposing user prompts. The attack needs only passive access to cache blocks and public model dictionaries.
Do this
- Check whether KV-Cloak or similar shuffled-obfuscation KV cache defenses are deployed, and track vendor guidance on the reported issue.
- Rotate row-level secret parameters on a shorter schedule so one calibration cannot carry across many requests.
- Restrict access to KV cache blocks and offloaded cache storage to the tenant that created them.
- Treat obfuscated KV caches as potentially revealing token content until a patched defense is confirmed.
Signals
- NoCISA KEV
- 23/100Priority
- 60%Confidence
Evidence
“Refreshing row-level secrets between epochs interrupts cross-request transfer, confirming secret reuse scope as the operative vulnerability.”
“achieves 100 percent token-multiset recovery on GPU”
Sources
Affected / related
Repeat-to-Reveal, KV-Cloak, PermLLM
Published 2026-11-18 11:53 UTC · updated 2026-11-19 04:35 UTC · AI-assisted summary, checked against quoted sources · How we score