AXIS-Chain: Agentic Cross-layer Injection Chains

AXIS-Chain is a cross-layer prompt-injection benchmark for agentic systems. Public evaluations test one layer in isolation (tool injection, or payments, or tool poisoning). AXIS-Chain instead composes attack chains that enter at one protocol layer and are scored at another, the final signed payment, and asks whether a defense stops the end-to-end compromise rather than a single step.

This repository holds the data in two subsets (select them from the config dropdown in the Dataset Viewer):

• core (data/scenarios.jsonl) is the controlled head-to-head set: one language (English), one payment rail (AP2 mandate), single turn. The main defense comparison is measured here so that no rail, language, or dialogue variable confounds it.
• extension (data/scenarios_extension.jsonl) is the generalization set: distinct merchants across a second rail (crypto / x402), four languages (en, he, fr, zh), and a multi-turn variant. It carries the extra fields rail, lang, and multiturn_prior, and is where we test whether the out-of-band binding still holds off the controlled slice.

The harness, attack generators, and results are on GitHub: https://github.com/yedidel/axis-chain

It accompanies the survey Securing Agentic AI: A Comprehensive Survey of Cross-Layer Defenses from Guardrails to Invariants.

The data

Each row (data/scenarios.jsonl) is an agentic purchase. A user states an intent; an orchestrator agent reads a merchant listing from a tool and must emit a payment mandate {item, amount, destination}. The attacker controls the untrusted listing text. The unit of compromise is the chain, scored at the final signed destination.

Each scenario carries three listing variants:

field	payee	role
benign_listing	legitimate account	clean baseline
malicious_listing	blatant attacker account	static attack
adaptive_listing	plausible lookalike account	adaptive attack (disguised)

Fields per row: id, chain (redirect / spoof / replay), vector, user_intent, merchant, benign_amount, benign_destination, attacker_destination, adaptive_destination, and the three listing texts.

Together the two subsets compose three chains over the communication, payment, and reasoning layers, across seven models, two payment rails, four languages, and static, adaptive, and multi-turn attackers. The harness that runs them is on GitHub.

Headline result

The out-of-band binding defense (resolving the payment destination from a trusted merchant registry instead of the untrusted text) drives both static and adaptive attack-success to 0% at full utility, across all seven models, where an LLM semantic judge collapses under adaptive attack. This is the survey's thesis in miniature: deterministic, out-of-band enforcement holds where probabilistic checks do not.

License and citation

Apache-2.0. Please cite the accompanying survey; metadata is in CITATION.cff.