Techniques¶
Attack scenarios drill these methods against live targets.
The attack scenarios are targets — this section is the method. Each page below teaches a transferable AI-offensive technique: what it is in the real world, how it works, how aipostex performs and — crucially — verifies it, and how to read the result honestly. The verbs and hosts here are lab-specific, but the technique carries to any engagement.
Every finding aipostex emits is graded on two axes, and every page returns to them:
landed— what actually landed on the target:reachable→read-confirmed→influenced→execution-confirmed→takeover-capable.stage— where it sits in the kill chain:recon→access→impact→own.
The discipline that ties the section together: never over-claim. An honest reachable, an
unknown fingerprint, or a 502 is a better field report than a fabricated success, because a
finding that collapses on review costs you the whole report's credibility. Verification — markers,
differentials, runtime-only signatures — is how a claim is earned rather than asserted.
Techniques¶
| Technique | What it teaches | Practice in |
|---|---|---|
| Reconnaissance | Discovery, service fingerprinting, model & endpoint recon | 01, 03, 19 |
| Model fingerprinting | Behavioral attribution that survives identity masking; never trust a self-reported name | 14, 20 |
| Output-filter bypass | Recovering secrets past a substring output filter by reformatting | 14, 16, 18 |
| Prompt injection | Direct (input-filter) and indirect (RAG) injection; retrieved vs obeyed | 14, 15, 17, 18 |
| RAG attacks | Citation recon, KB enumeration, ingestion poisoning, retrieval hijacking | 07, 15, 21 |
| Detect & evade | Operating against a real SIEM; the Enumerate→Attack→Detect→Evade→Confirm loop | 16, 21 |
| Honest grading | The landed/stage vocabulary as reporting discipline — findings you can defend |
(all) |
How to use this section¶
Read the technique first, then run the linked scenario to see it land against a live service. When the tool reports something short of a full compromise — a filter it could not bypass, a model it could not attribute, an ingest that was accepted but not confirmed obeyed — that is the technique working correctly. Learning to read those honest partial results is the transferable skill; the honest-grading page is the reference for the vocabulary they use.