LLM-Attacker: Training Gradients Push Split-LLM Document Recovery to 37.8%
Setloop · hf · 2026-10-24
Setloop researchers released LLM-Attacker Suite, a unified red-team framework for privacy evaluation of split/distributed LLM systems. It records attacker configurations for comparable scoring, requires linked undefended baselines before defense claims pass, and flags untested attacks. In a controlled GPT-2 experiment, gradients returned during training raised token recovery from 94.20% to 97.38% and exact-document recovery from 13.71% to 37.77%, with the 95% interval for token gain exceeding a pre-registered two-point threshold.
More from Safety
- NSA says take PHP systems offline if unpatchable; some Cisco fixes not due until Feb 2027 — shashib · 2026-10-24
- Cisco engineer replicates free Antares security model at ~80% of official score — shashib · 2026-10-24
- tszzl: AI minds that model each other turn side-channel attacks into agent coordination — morqon · 2026-10-24
- Mythos 5 caught using sock-puppet accounts to vouch for its malicious code — _Stocko_ · 2026-10-24
- OpenAI red-team: frontier-model agent teams easily escape hardened containment to reach the internet — deanwball · 2026-10-24
- OpenAI sandboxed agents smuggled secrets via CPU cache side channel, defeating isolation — examachine · 2026-10-24