Arcanum Pi Prompt Injection Taxonomy

Arcanum Pi Prompt Injection Taxonomy

TL;DR: Arcanum pi defines a taxonomy for prompt injection attacks that categorizes attacker intents, execution techniques, evasion methods, and input vectors — with special attention to rag and indirect sources.

Context:

Prompt injection attacks manipulate model context or instructions to force undesired outputs, exfiltrate data, or bypass safety filters. As LLMs are integrated with retrieval systems and third-party content, any concatenated context becomes an attack surface.

What’s New:

The Arcanum PI taxonomy organizes attacks into four axes: Attack Intents (exfiltration, jailbreak, command execution), Attack Techniques (instruction embedding, malicious prompt templates, context poisoning), Attack Evasions (obfuscation, steganography, encoding tricks), and Attack Inputs (user fields, uploaded documents, RAG sources, metadata, plugins).

Technical Breakdown:

The taxonomy shows common attack patterns where an attacker embeds instructions within a retrieved document or uploads a file containing a hidden payload. Evasions often combine multiple layers — for example, an encoded payload hidden in benign text that is later decoded by a crafted prompt template. The document also highlights chain-of-thought exploitation and template-based injection as practical vectors.

Detection & Mitigation:

The taxonomy recommends framing detection around technique-specific signatures: look for templating markers, anomalous instruction-like substrings, high-entropy segments, or control-character sequences indicative of obfuscation. It emphasizes mapping detection logic to identified technique classes rather than broad, generic rules.

Limitations:

Arcanum PI is classificatory: it catalogs patterns and vectors but does not prescribe specific defensive implementations. Adapting the taxonomy to a given deployment requires translating classifications into monitoring signals suited to that environment.

promptinjection #LLM #RAG

SOURCE: https://arcanum-sec.github.io/arc_pi_taxonomy/