Last updated
What's in the checklist?
Every category of the OWASP Top 10 for LLM Applications, restated as plain-language assessment questions and mapped to the corresponding expectations in the Australian Government AI Technical Standard. So one pass over your system answers both "how would this be attacked?" and "what will government reviewers ask?".
- LLM01 Prompt injection. Direct and indirect injection paths, and what untrusted content can reach the context window.
- LLM02 Sensitive information disclosure. System prompts, RAG sources and cross-tenant leakage.
- LLM03 Supply chain. Model provenance, third-party components and fine-tuning data.
- LLM04 Data and model poisoning. Ingest pipelines and knowledge-base integrity.
- LLM05 Improper output handling. Everything downstream code trusts a model to say.
- LLM06 Excessive agency. Tools, permissions and what the model can actually do.
- LLM07 System prompt leakage. What your instructions reveal and to whom.
- LLM08 Vector and embedding weaknesses. Retrieval isolation and embedding-layer attacks.
- LLM09 Misinformation. Accuracy controls where outputs carry consequences.
- LLM10 Unbounded consumption. Rate limits, cost controls and denial-of-wallet.
- The crosswalk. Each item mapped to the AU AI Technical Standard's security, testing, transparency and monitoring statements, with a column for evidence.
Who is this for?
Security and engineering leads assessing an AI feature before launch, teams preparing for a government or enterprise review, and anyone who has been asked "is our AI secure?" and would prefer to answer with a document rather than a shrug. If you'd rather have the assessment done for you, that's an LLM penetration test or a broader AI security engagement.
About Ironbark Cyber
Ironbark Cyber is an Australian AI security and penetration testing consultancy, founded in 2025 by Luke Stephens (hakluke) and part of the Haksec group. This checklist is the one our consultants start from on real engagements. Lightly sanitised, otherwise unchanged.
Download
Get the checklist
Tell us who you are and we'll send it over. That's the whole transaction.