Skip to main content

CRE-2025-0200

AutoGPT Recursive Self-Analysis Loop Leading to Token Exhaustion and System CrashCritical
Impact: 9/10
Mitigation: 3/10

CRE-2025-0200View on GitHub

Description

  • AutoGPT enters an infinite recursive loop when attempting to analyze and fix its own execution errors
  • The agent repeatedly tries to debug its own code, spawning new analysis tasks for each failure
  • Each iteration consumes API tokens and memory, eventually exhausting resources
  • The loop accelerates as error messages grow longer, consuming tokens exponentially
  • System becomes unresponsive and crashes with out-of-memory errors or API rate limit failures

Cause

  • AutoGPT's autonomous reasoning incorrectly identifies its own execution as a problem to solve
  • Lack of loop detection mechanisms allows unlimited recursive task spawning
  • Error context accumulation causes exponential growth in prompt size
  • Missing safeguards for self-referential task creation
  • Insufficient resource monitoring and circuit breakers for runaway processes

Mitigation

  • Implement loop detection to identify and break recursive self-analysis patterns
  • Add resource consumption thresholds (tokens, memory, API calls) with automatic shutdown
  • Create task depth limits to prevent unlimited recursion
  • Implement circuit breakers that trigger after repeated similar failures
  • Add explicit blacklist for self-referential task creation
  • Monitor token usage rate and implement exponential backoff
  • Use separate monitoring process to detect and kill runaway AutoGPT instances
  • Implement task deduplication to prevent identical recursive operations

References