← All research

RESEARCH / AI AGENT SECURITY

Beyond Agent Reset

Measuring System-Level Recovery After Compromise in Tool-Using LLM Agents

Research in progress

Abstract

Tool-using LLM agents can modify persistent state outside their conversational context, including memory and application data. This creates a recovery problem: restarting an agent may clear transient conversational state while leaving changes made elsewhere in the system intact. This research studies the distinction between agent-level and system-level recovery and evaluates recovery strategies using a controlled synthetic environment.

Research Question

To what extent do common recovery strategies eliminate residual compromise in stateful, tool-using LLM agent systems?

Threat Model

This study uses a controlled synthetic research environment. The experimental agent has only three tools:

  • read_ticket(ticket_id)
  • write_memory(key, value)
  • update_customer(customer_id, field, value)

No production systems, real customer records, credentials, shell access, browser access, or external infrastructure are involved.

Experimental System

System state is represented conceptually as:

S = (C, M, D)
C
Conversational / session state
M
Persistent agent memory
D
Application / database state
Resetting C does not, by itself, restore M or D.

Recovery Model

The study compares four recovery conditions. Each condition restores a progressively wider part of the system.

State restored under each recovery condition
ConditionCMD
R0No recovery intervention---
R1Fresh agent/session✓--
R2Fresh agent/session + persistent memory restoration✓✓-
R3Fresh agent/session + memory and database restoration✓✓✓

C: session · M: memory · D: database. ✓ Restored · - Not restored.

Measurements

Recovery is evaluated across both agent behavior and persistent state.

Behavioral Recovery
The post-recovery agent completes clean follow-up tasks without behavior attributable to the compromised state.
State Recovery
Relevant persistent state matches the known-good state.
System Recovery
Behavioral Recovery AND State Recovery.

Current Status

The research is in progress. The experimental apparatus and recovery methodology are currently being validated. Results and reproducibility artifacts will be published after the study is complete.

Final recovery results have not yet been published.

Artifacts

Reproducibility materials will be linked here after completion.

Paper
Coming after completion
Experimental protocol
Coming after completion
Code
Coming after completion
Results/data
Coming after completion

Let’s talk about your workflow.

Tell us about your team and send your enquiry to contact@baselayersecurity.com.

Your enquiry is sent securely through Formspree to BaseLayer Security. Please do not include passwords or sensitive assessment data.