Skip to content

Failure Modes

What AgentBreak simulates — and what's out of scope.

Targets

  • llm_chat — OpenAI (/v1/chat/completions) and Anthropic (/v1/messages) chat APIs
  • mcp_tool — MCP tool calls, resource reads, prompt gets

Fault support matrix

Fault LLM MCP Notes
http_error yes yes Requires status_code
latency yes yes Requires min_ms, max_ms
timeout no yes Requires min_ms, max_ms; sleeps then returns 504
empty_response yes yes
invalid_json yes yes
schema_violation yes yes Corrupts tool_calls (LLM) or result shape (MCP)
wrong_content yes yes Optional body field
large_response yes yes Requires size_bytes > 0

Built-in behavioral detection

  • Duplicate request detection — fingerprint-based, for both LLM and MCP
  • Suspected loop detection — 3+ identical requests
  • Expired upstream MCP session recovery — re-initializes on session errors

Out of scope (for now)

  • Prompt injection
  • Memory poisoning
  • Queue replay
  • Checkpoint corruption
  • Browser worker failures
  • Multi-agent coordination failures

See Roadmap for planned future targets.