Topics
Browse everything Prompt Injection Report publishes by category and tag: every topic covered here, with the most recent guides listed under each one.
Tags
- #prompt-injection 13
- #llm-security 9
- #indirect-prompt-injection 7
- #agentic-ai 4
- #owasp-llm 3
- #red-team 3
- #defense 2
- #detection 2
- #garak 2
- #guardrails 2
- #pyrit 2
- #rag 2
- #testing 2
- #threat-modeling 2
- #agent-security 1
- #agentdojo 1
- #ai-agents 1
- #ascii-smuggling 1
- #benchmarks 1
- #bing 1
- #bypass 1
- #canary-tokens 1
- #case-study 1
- #ci 1
- #compliance 1
- #eu-ai-act 1
- #incident-analysis 1
- #invisible-text 1
- #llama 1
- #mitigations 1
- #multi-modal 1
- #open-source-models 1
- #policy 1
- #promptmap 1
- #real-world 1
- #rebuff 1
- #regulatory-liability 1
- #sydney 1
- #taxonomy 1
- #tokenization 1
- #tool-abuse 1
- #tooling 1
- #unicode 1
Categories
Defense 3 posts
- Prompt Injection Detection Techniques That Actually WorkCompare prompt injection detection techniques: pattern filters, fine-tuned classifiers, and known-answer probes, plus why no single layer closes the hole.
- How to Mitigate Prompt Injection: A Layered Defense GuideNo single control stops prompt injection. This guide covers the layered stack that works: input and output filtering, privilege limits, and isolation.
- Rebuff Prompt Injection: Architecture, Limits, and GapsRebuff uses four prompt injection checks: heuristics, LLM classification, vector similarity, and canary tokens. This review maps their blind spots.
Offensive 2 posts
- Invisible Prompt Injection via Unicode Tag SmugglingUnicode Tag characters let attackers embed invisible prompt injection payloads that still tokenize as instructions. How it works and what stops it.
- Llama 3 RAG Prompt Injection: Payloads and DefensesA documentation-based walkthrough of four indirect prompt injection payloads for a Llama 3.1 RAG architecture and the published defenses that limit them.
Primer 2 posts
- Indirect Prompt Injection: How These Attacks WorkIndirect prompt injection hides instructions in content the model retrieves. The delivery channels, the attacker objectives, and what benchmarks measure.
- How Prompt Injection Attacks Work: A Working TaxonomyPrompt injection covers five classes: direct, indirect, multi-modal, agentic, and multi-turn, each with its own trust boundary and its own working defense.
Tooling 2 posts
- Prompt Injection Testing: A Repeatable MethodPrompt injection testing needs a method, not just a scanner. How to scope surfaces, pick a corpus and harness, define detectors, and score attack success.
- Garak vs PyRIT vs promptmap: Prompt Injection TestingGarak, PyRIT, and promptmap are not interchangeable. What each prompt injection testing framework is built for, what it misses, and how to choose one.