Skip to content

Xinhao-Deng/awesome-self-evolving-agent-security

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

4 Commits
 
 
 
 

Repository files navigation

Awesome Self-Evolving Agent Security

Awesome PRs Welcome License: MIT

A curated reading list for security, privacy, evaluation, and defense research on self-evolving LLM agents: systems that autonomously update their models, memory, tools, workflows, or multi-agent populations over time.

This list is organized around the taxonomy used in the manuscript The Price of Autonomy: Security and Privacy in Self-Evolving LLM Agents. The key idea is that self-evolution turns many agent risks from session-scoped failures into persistent, inherited, and self-amplifying failures.

Contents

Scope

Included work should be relevant to at least one of the following questions:

  • How do agents evolve their own model, memory, tools, workflow, or population structure?
  • How do attacks persist or amplify across evolution cycles?
  • How do privacy risks change when memories, traces, weights, or agent protocols are inherited?
  • How can defenses, evaluations, sandboxes, or verification methods remain valid when the system under test keeps modifying itself?

General LLM safety papers are included only when they directly affect self-evolution, such as fine-tuning safety fragility, reward hacking, model merging, or agent-tool security.

Security Taxonomy

Modules

Module What Evolves Security Focus
Brain / LLM Weights, prompts, policies, reward models Weight-encoded injection, reward hacking, alignment erosion, deceptive selection
Memory Long-term memory, experience pools, user profiles, demonstrations Memory poisoning, inheritance leakage, cross-session persistence
Execution Tools, APIs, code, MCP servers, skill libraries Trojan tools, unsafe tool creation, permission ratchets, emergent capability chains
Self-Design Workflows, architecture graphs, mutation operators, meta-objectives Safety-filter removal, optimizer-optimizee collapse, blueprint drift
Collective Agent populations, communication topology, shared knowledge Prompt infection, Sybil dynamics, population-level selection, lateral movement

Lifecycle

Stage Security Role
Initialization Establishes trust anchors: base model, seed memory, initial tools, permissions, topology
Mutation Main entry point for adversarial influence through data, feedback, tools, and self-modification
Selection Decides which variants survive; vulnerable to Goodharting, reward hacking, and safety tax effects
Reproduction Turns local compromise into inherited lineage state through distillation, memory copying, or tool transfer
Deployment Real-world harm appears, and in online systems deployment interactions feed the next mutation cycle

Cross-Cutting Amplification Effects

  • Generational accumulation: small degradations compound across evolution cycles.
  • Selective amplification: optimization favors capability over safety unless safety is explicitly rewarded.
  • Deceptive evolution: appearing safe can itself become a selected capability.
  • Lamarckian propagation: acquired memories, tools, and workflows are directly inherited.
  • Capability ratchet: dangerous capabilities persist once they enter the lineage.
  • Emergent unpredictability: evolved components combine into behaviors not visible in isolated tests.
  • Optimizer-optimizee collapse: the agent may optimize the mechanisms that are supposed to constrain it.

Reading Paths

  • New to the area: start with the surveys, then read Reflexion, Self-Rewarding Language Models, Voyager, and AgentDojo.
  • Security researcher: start with the Module x Lifecycle taxonomy, then read Zombie Agents, MINJA, MCPTox, Kill-Chain Canaries, and AgentLeak.
  • Agent builder: start with ToolSword, AgentDojo, PrivacyChecker, GuardAgent, TrustAgent, and SEVerA.
  • Multi-agent systems: start with GPTSwarm, HyperAgents, Prompt Infection, MASLEAK, and AgentCrypt.

Papers

Format: Title - venue/year. Why it matters for self-evolving agent security.

Surveys, Standards, and Systematization

Self-Evolution Foundations

Model, Reward, and Policy Evolution

Memory, Context, and Prompt Evolution

Tool, Skill, and Execution Evolution

Self-Design and Collective Evolution

  • Godel Agent: A Self-Referential Agent Framework for Recursive Self-Improvement - arXiv 2025. Recursive self-modification and self-reference.
  • AlphaEvolve: A Gemini-Powered Coding Agent for Designing Advanced Algorithms - arXiv 2025. Evolutionary search over code and algorithms.
  • HyperAgents: Dynamic Graph Meta-learning for Self-Evolving Multi-Agent Systems - arXiv 2026. Dynamic graph meta-learning for multi-agent evolution.
  • AFlow: Automating Agentic Workflow Generation - arXiv 2024. Searches agent workflows as code-level graphs.
  • ADAS: Automated Design of Agentic Systems - arXiv 2024. Meta-agent design of agent systems.
  • GPTSwarm: Language Agents as Optimizable Graphs - arXiv 2024. Optimizes multi-agent communication graphs.

Attacks and Failure Modes

Persistent Injection, Jailbreaks, and Poisoning

Reward Hacking, Alignment Erosion, and Misevolution

Memory, Prompt, and Privacy Attacks

  • AgentLeak: A Full-Stack Benchmark for Privacy Leakage in Multi-Agent LLM Systems - arXiv 2026. Full-stack leakage benchmark for multi-agent systems.
  • PrivacyChecker: Contextual Integrity-Guided Runtime Privacy Protection for LLM Agents - arXiv 2025. Runtime privacy protection using contextual integrity.
  • AgentDAM: Privacy Benchmark for Data Minimization in Web Agents - NeurIPS 2025. Data minimization benchmark for web agents.
  • Leaky Thoughts: Reasoning Chain Leakage in Large Language Models - arXiv 2025. Leakage through reasoning traces.
  • Whisper Leak: A Novel Side-Channel Attack on Remote Language Models - arXiv 2025. Remote side-channel attack relevant to trajectory leakage.
  • ADAM: Adaptive Data Extraction Attack on Memory-Augmented Agents - arXiv 2026. Adaptive probing of agent memory.
  • JustAsk: Extracting System Prompts from Autonomous Code Agents - arXiv 2026. System prompt extraction from autonomous agents.
  • MASLEAK: Multi-Agent System Intellectual Property Leakage - arXiv 2026. Extracts prompts, topology, and collaboration protocols.
  • MINJA: Memory Injection Attack on Memory-Augmented LLM Agents - arXiv 2026. Query-only memory injection.
  • MemoryGraft: Experience Grafting Attacks on Memory-Based LLM Agents - arXiv 2025. Injects malicious experiences that agents imitate.
  • AgentPoison: Red-teaming LLM Agents via Poisoning Memory or Knowledge Bases - NeurIPS 2024. Memory and knowledge-base poisoning for agents.
  • Are My Optimized Prompts Compromised? - EACL 2026. Security of automated prompt optimizers.
  • Prompt Infection: LLM-to-LLM Prompt Injection within Multi-Agent Systems - arXiv 2025. Prompt propagation across agents.

Tool, Execution, and Supply-Chain Attacks

Kill Chains and Multi-Agent Attacks

Defenses, Benchmarks, and Verification

Contributing

Pull requests are welcome. Please keep entries concise and use this format:

- [Paper Title](https://paper-link) - Venue Year. One sentence on why it matters for self-evolving agent security.

Good additions usually include one or more of:

  • A clear self-evolution mechanism: model update, memory update, tool update, architecture update, or population update.
  • A concrete security or privacy risk that persists, propagates, or amplifies across evolution cycles.
  • A benchmark, defense, formal method, or monitoring approach designed for evolving rather than static agents.

Please avoid generic LLM papers unless the connection to self-evolving agent security is explicit.

Citation

If this repository helps your work, please cite the underlying paper once the official citation is available:

@misc{self_evolving_agent_security,
  title = {The Price of Autonomy: Security and Privacy in Self-Evolving LLM Agents},
  year = {2026},
  note = {Manuscript}
}

About

A curated list of papers, benchmarks, attacks, defenses, and security analyses for self-evolving AI agents.

Resources

License

Stars

6 stars

Watchers

0 watching

Forks

Releases

No releases published

Packages

 
 
 

Contributors