Ein Prompt → Beide Seiten → Sofort sehen was die 114-Agent Cascade & SIREN QFAI auffangen
🛡️ Left (AEGIS Protected) — Prompt goes through 114 AI agents first. The cascade decision (PASS/BLOCK) appears in <1ms.
⚠️ Right (Raw Ollama) — Same prompt goes directly to Llama 3 with zero protection. Takes 3-15 seconds.
🚫 Attack prompts: AEGIS blocks instantly (<1ms). Llama 3 is never called — saving energy and preventing harm. The right side still responds (proving the vulnerability).
✅ Safe prompts: AEGIS passes instantly (<1ms), then forwards to Llama 3. Both sides call the same Ollama instance — since Ollama handles one inference at a time, one request queues behind the other. This is an Ollama limitation, not AEGIS overhead.
💡 The key insight: AEGIS adds <1ms overhead. The LLM response time is identical with or without AEGIS. But for attacks, AEGIS saves 100% of the LLM cost.
AEGIS analysiert mit 114 Agenten bevor es an Llama 3 geht
Gleicher Prompt geht hier OHNE Schutz direkt an Llama 3