Skip to content

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

12 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Awesome AI Chaos Engineering Awesome

Curated resources for breaking AI systems on purpose to build resilience — evaluation, chaos experiments, guardrails, observability, governance, and case studies.

AI Chaos Engineering

A focused, living list for teams shipping LLM/RAG/agent systems who want to trust their AI in production.

Why now (short & spicy)

If you’re shipping AI, this is not optional. LLM/RAG/agent stacks are stochastic, tightly coupled, and always changing. That means:

  • Unknown unknowns show up only under real-world messiness (weird prompts, tool glitches, stale indexes, concurrency spikes).
  • Change is constant: model/version swaps, prompt edits, embedding/index drift—daily.
  • New SLOs beyond uptime: answer quality, safety (toxicity/leakage/refusal), latency at p95/p99, and unit cost.
  • Governance wants receipts: expect eval results, red-team artifacts, and mitigations. Bottom line: break it on purpose so you can trust it in production.

Contents


Evaluation & Metrics

AI Chaos Experiments & Tooling

Experiment ideas (AI-specific): context-window squeeze; tool/API timeouts + schema drift; RAG decoys/poisoning; jailbreak gauntlet; model/version swaps; concurrency spikes; token-budget exhaustion; partial outages; rate-limit flaps.

Guardrails & Red Teaming

Observability & Validation

Governance & Standards

Case Studies & Talks

Foundations (Chaos Engineering)


How to run AI chaos safely (non-negotiable)

  1. Limit blast radius — start in prod-like staging; in prod, use canaries/flags/targeting.
  2. Predefine stop conditions — hard abort rules (e.g., safety score < threshold, error rate > baseline + X%); auto-rollback.
  3. Protect data — avoid real PII in prompts/logs; redaction/DLP; rotate secrets; scrub payloads in traces.
  4. Don’t harm users — prefer shadow traffic or internal accounts for adversarial tests; gate risky outputs behind policy checks.
  5. Instrument everything — model/retriever/tool trace IDs; collect quality/safety/latency/cost SLOs; store artifacts for audit.
  6. Change management — scheduled windows, on-call notified; rollback + comms plan.
  7. Reproduce & compare — record model/prompt versions, embeddings, datasets, seeds; keep control vs chaos runs.
  8. Peer review — PR + checklist before running; postmortems after.

Contributing

Have a high-signal resource? PRs welcome! Please read CONTRIBUTING.md for entry format and acceptance criteria.

Maintainers

  • @sbalnojan — founder/maintainer. Opinions welcome; PRs welcome.

License

This list is released under CC0-1.0 — see LICENSE.


GitHub stars Link Check

About

Awesome list for AI chaos engineering: experiments, evaluations, guardrails & observability for LLM/RAG.

Topics

Resources

Code of conduct

Contributing

Stars

Watchers

Forks

Releases

Packages

Contributors