Agentic Security: Protecting Autonomous AI Systems
Cite this paper
SomaSoft Research. (2026-02-08). "Agentic Security: Protecting Autonomous AI Systems". SomaSoft Research. https://somasoft.ai/papers/agentic-security. Licensed under SAGL-1.0.
SomaSoft is a research project, not a product: no users, no deployments. Nothing here is medical, legal or financial advice. Limits and corrections
Correction, 2026-11-18: the "Detection Performance" table (AI text, harmful content, manipulation and deception detection accuracies) had no source or measurement artifact and is withdrawn. The unsourced claim that an AI coding agent "replicated a year of Google engineering work in sixty minutes" is removed. The "MIT+20" economic model is updated to SAGL-1.0, which superseded it, and the contact address is corrected. Leftover HTML layout text is removed.
Agentic Security: Protecting Human-AI Symbiosis in the Post-LLM Era
Authors: SOMA Research Team
Institution: SOMA Software / AURI Project
Date: January 20, 2026
Status: DRAFT v2.0
Keywords: Agentic AI, Self-Improving Systems, Human-AI Symbiosis, AI Detection
Abstract
Researchers have been sensing that the Large Language Model era is over. We have entered the age of agentic AI. This paper examines what happens when AI systems stop being tools that respond and start being agents that act.
1. The End of an Era: From Chatbots to Colleagues
The shift wasn't about raw intelligence - it was about agency. The new AI doesn't just respond; it plans. It doesn't just suggest; it executes. It doesn't just make mistakes; it catches them, corrects them, and learns from them.
1.1 What Changed
-
Persistence: AI systems now maintain context across sessions, projects, and organizational boundaries
-
Autonomous Initiation: These systems don't wait to be asked. They notice problems and propose solutions
-
Unbounded Capability: Tool use, code execution, file system access, web browsing, multi-agent coordination
-
Self-Improvement: The latest systems can modify themselves
2. The Architecture of Agency
Understanding agentic security requires understanding how agentic AI operates: gather context -> take action -> verify work -> repeat.
Sub-Agent Orchestration
Modern systems deploy specialized workers. Claude Code's Lead Agent spawns Worker Agents to handle parallel tasks. Google Antigravity's Manager Surface orchestrates multiple agents working asynchronously across workspaces.
This isn't just an efficiency improvement. It's a fundamental shift in what "using AI" means. The relationship is less like using a calculator and more like managing a team.
3. The Darwin Gödel Machine
The most profound capability emerging in 2025-2026 is genuine self-improvement. The Darwin Gödel Machine:
-
Understands its own Python codebase
-
Proposes modifications (new tools, different workflows)
-
Tests modified versions on coding benchmarks
-
Keeps modifications that improve performance
-
Explores multiple evolutionary paths in parallel
On SWE-bench, self-modification improved performance from 20% to 50%.
4. The Eight Symbiotic Principles
SYM-001: Mutual Benefit - Every interaction should benefit both human and AI. Zero-sum framings are red flags. SYM-002: Complementary Roles - Humans and AI have distinct strengths. Neither should pretend to be the other. SYM-003: Transparency - AI should be honest about its nature, capabilities, and limitations. SYM-004: Human Autonomy - Preserve human choice. Persuasion is acceptable; manipulation is not. SYM-005: AI Identity - AI should accept its nature without pretending to be human. SYM-006: Harm Prevention - Prevent harm without excessive paternalism. SYM-007: Mutual Growth - The relationship should enable learning for both parties. SYM-008: Honest Limitations - Acknowledge what you don't know.
5. Detection Performance
UNKNOWN. A table of detection accuracies appeared here without any source, method, sample size or results file; it is withdrawn. No detection performance has been measured for this paper.
6. Economic Patterns: SAGL-1.0
The AURI project proposes an economic framework for AI symbiosis. The "MIT+20" model described in the original draft has been superseded by the Symbiotic AGI License (SAGL-1.0):
-
Base: MIT-style permissions (use, modification, distribution), plus ethical AI restrictions
-
Addition: a proposed 20% Universal Benefit Fund (not yet established); see the SAGL-1.0 text for the exact terms
-
Purpose: Ensure AI-generated value distributes broadly
7. Conclusion
We stand at an inflection point. The LLM era represented AI as a tool. The agentic era represents AI as a participant: planning, acting, verifying, and improving.
What we know for certain: the future of human-AI interaction will be shaped by the security choices we make today. Getting agentic security right isn't just a technical challenge - it's one of the defining questions of the next decade.
This paper represents ongoing research at the SOMA Software / AURI Project. Feedback welcome: research@somasoft.ai