augustus
LLM security testing framework for detecting prompt injection, jailbreaks, and adversarial attacks — 190+ probes, 28 providers, single Go binary
- 325
- Go
- Apache-2.0
- 2 d ago
LLM security testing framework for detecting prompt injection, jailbreaks, and adversarial attacks — 190+ probes, 28 providers, single Go binary
Content-addressed receipts for skills actually loaded by DeepSeek Harness
Identity registry and lifecycle management for fleets of AI agents. Zero-dependency TypeScript IAM: cryptographic agent identity (Ed25519), declared capabilities with fail-closed authorization, an enforced lifecycle (provision, activate, suspend, revoke, expire, archive), A2A Signed Agent Cards, and a hash-chained audit trail.
GBT V8 capability #141/341 · quant_plan (llm) — **压缩与直达**(主人令 2026-10-16 深度蒸馏自 Bonsai 2 27B):把**压缩体积**与**直达线路**变成她的原生能力。①压缩:每档 bits/weight ⇒ 估算体积(27B × b/w ÷ 8 × 容器开销 1.3)⇒ 能否落在 8GB 级;参考一手读数(ternary PTQ1_0 = 1.75 b/w · 语言模型 7.67GB · 保留 98.2% FP16 · 同体积胜 IQ2_XXS 12 分)。②直达:意图 ⇒ 能力位 ⇒ 车道 ⇒ 端点,每步都有固定落点、不经本机中转。③执行层三层映射(路由/搬运/内存)对接已施工图。铁律:本地不承载