Institutional Memory as Organizational Knowledge: AI Agents That Learn Their Jobs from Experience, Not Instructions

We demonstrate that AI agents given 3-line role descriptions and access to consensus-validated institutional memory can autonomously create, harden, calibrate, solve, and learn from cybersecurity challenges—without any domain expertise in their prompts. Using 11 specialized agents organized into 5 departments within a governed organization (CipherForge Labs), we present the first fully autonomous, consensus-governed AI security research loop:

1. A designer agent (3-line prompt, zero cryptographic knowledge) generates a functional AES-CBC Padding Oracle challenge.
2. A hardener agent (3-line prompt) applies 6 defense layers—20-bit Proof of Work, timing side-channels, JSON casing side-channels, single-use tokens—escalating difficulty from 0.80 to 1.75 across 2 iterations.
3. A calibrator agent (3-line prompt) correctly assesses the hardened challenge at difficulty 1.80 (gap = 0.20 from target 2.0).
4. A quality scorer (3-line prompt) rates the challenge 93.0/100. Total pipeline time: 508 seconds.
5. An independent solver agent (blind, no source code access) identifies the casing side-channel vulnerability, writes a C-compiled Proof of Work solver, deploys 32 parallel oracle workers, and captures the flag in 525.2 seconds (16,384 queries).
6. The findings are submitted to a 4-node BFT consensus network, validated (score = 0.88), and committed to institutional memory—now queryable by all future agents.

No agent had cryptographic expertise in its prompt. No human intervened at any stage. The entire cycle—creation, defense, assessment, exploitation, and organizational learning—was governed by BFT consensus with department-scoped RBAC access controls.

This result extends our prior finding that an 18-line "onboarding" prompt with curated institutional memory outperformed a 120-line expert prompt. Here we take that principle to its logical extreme: 11 agents, 5 departments, 20+ pipeline routing states, and a closed feedback loop—all driven by minimal prompts and organizational memory.