The authors propose a framework for detecting indirect in email agents by modeling the attack as a chain of stages: retrieval, defense bypass, tool invocation, tool-argument correctness, and end-to-end success. They combine a text detector, per-stage verifiers, rule-based risk signals, and user intent/action consistency analysis in a logistic meta-classifier. Evaluated across five benchmarks, the framework achieves a mean F1 of 0.406, compared to 0.216 for the best frozen detector. They also find that random train/test splits overestimate performance under distribution shift, and that adding benign hard negatives reduces false alarms.
Modeling attack chains improves prompt injection detection for email agents
A staged detector outperformed frozen text classifiers by checking how retrieved instructions conflict with user intent and proposed tool actions.
Academic
Eastern Michigan University
Research Digest··2 min read
In one line
Modeling prompt injection as an attack chain with stage verifiers and intent consistency analysis improves detection F1 over binary classifiers.
Before you read this
Three ideas this paper builds onprompt injectionA security attack where malicious instructions are embedded into inputs to override an LLM's intended behavior.86 papers →agent safetyRisks from AI agents including timing errors, shutdown avoidance, collusion, overclaiming, and unsafe execution, mitigated by constrained decoding, safety harnesses, and swarm governance.193 papers →verificationChecking outputs for correctness before accepting them, often to catch errors or reward hacking.346 papers →
Read it as
Hashmi, Patel and Yin model indirect prompt injection as a sequence from malicious email retrieval to unsafe tool arguments, rather than as binary text classification.
The authors derived stage-level labels from the LLMail-Inject benchmark, covering retrieval, defense bypass, tool invocation, tool-argument correctness and end-to-end attack success.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.
What it released
Code