ACTS - Blind Cipher Identification: leaderboard
ACTS (Ibrahim and Salama, 2026) tests whether language models can tell which encryption algorithm produced a ciphertext file. This board is the blind setting, where models see only the ciphertext of 127 files from seven algorithm families with no filename or metadata; accuracy in percent with chance at about 14 percent, higher is better.
Metric: Accuracy (%; tier-3 blind identification of the algorithm family from ciphertext alone, no filename or metadata: 127 independently generated files, about 18 per family, in seven families (AES-128-CBC, AES-256-CBC, DES-CBC, 3DES, ChaCha20-Poly1305, RSA-OAEP-2048, ML-KEM-768) with unique random keys, plaintexts and padding lengths and no identifying headers; temperature 0, at most 512 output tokens; chance 14.3). Source: arxiv.org. Saturation forecast: Around September 2027. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Nemotron 3 Super | 37.3 | #766 |
| 2 | GPT-OSS-120B | 30.7 | #699 |
| 3 | Gemma 4 31B | 24.4 | #393 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=acts-blind-cipher-identification · How It Works · Data refreshed daily, snapshot 2026-11-12.