# On AgentDojo, a taxonomy-guided attacker finds more kinds of prompt injection

A new red-teaming method for AI agents tries every catalogued way of planting hidden instructions. It finds more distinct weaknesses than three recent attackers, though trained attackers still break more scenarios.

- Category: Safety & evaluation
- Readiness level: III (Open source) on Science Underway's four-step scale, as proposed
- Published: 2026-11-19
- Editor: Stefan Perdal
- URL: https://scienceunderway.com/articles/on-agentdojo-a-taxonomy-guided-attacker-finds-more-kinds-of-prompt-injection

## The study

AutoPI: Taxonomy-Guided Adaptive Red Teaming for Prompt Injection in LLM Agents (preprint, not peer reviewed). arXiv 2611.24313. https://arxiv.org/abs/2611.24313

## At a glance

- What it shows: An automated attacker that works through a fixed catalogue of attack styles finds working attacks in more distinct styles than three recent rivals. It also costs the least per attempt and needs no training.
- What it does not show: It does not show that models trained on these attacks become safer, or what the risk is in real use. One rival trained in advance still breaks more test scenarios.
- Why it matters: Security tests that only count how often an AI agent breaks can miss whole families of weakness. Counting how many different ways it breaks gives defenders a fuller picture.

## What the study shows

AI agents read emails, web pages and tool outputs that neither the user nor the developer wrote. Prompt injection hides instructions in that content, so the agent carries out the attacker's wishes. Red teaming, where testers attack a system to find its weak points, is the main way to measure this risk.

Most automated red teamers aim to maximise attack success rate (ASR), the share of test scenarios they manage to break. The authors argue this hides something. An attacker can score well by reusing one or two tricks, leaving other techniques untested.

Their method, AutoPI, works from PIAtlas, a catalogue of 132 injection sub-techniques grouped into 34 techniques and six classes. The same authors published the catalogue earlier. For each scenario, an attacker model writes one short injection per sub-technique. A second model then diagnoses failed attempts, and the attacker reworks them. The method only sees the target's outputs, needs no training and plugs into existing benchmarks.

The authors tested it on AgentDojo, a standard agent benchmark, and IPI Arena, a newer one used in vendor reports. The targets were the open model Qwen3.6-35B and the hosted Claude Haiku 5.5. They compared it with three published attackers (PIMiner, Climbing the Hill and AutoDojo) and with AgentDojo's own fixed attack.

The authors report that on AgentDojo against Qwen3.6-35B, AutoPI's successful attacks covered 31 of 34 techniques. The best rival covered 22. A separate AutoPI setting, which reworks AgentDojo's fixed attack, broke about 68% of scenarios at equal compute. That beat PIMiner but trailed Climbing the Hill, at about 84%. Both of those rivals are trained before the attack; AutoPI is not.

Against Claude Haiku 5.5, every attacker broke only four to six of 97 scenarios. AutoPI broke six. On IPI Arena it broke 14 of 41, slightly fewer than the two trained rivals. This is a preprint and has not been peer reviewed.

## Our reading

In our reading, the useful idea here is the yardstick more than the attacker. A security score that says an agent resists 90% of attacks says little if those attacks all use the same trick. Counting how many distinct techniques still work is a more honest summary of a worst-case risk.

The evidence for broader coverage on AgentDojo is reasonably solid. The lead held when attackers were compared at the same number of successes, and every attacker's output was labelled by the same procedure. The evidence for higher success rates is mixed. AutoPI leads only on Haiku, where the margins are a scenario or two, and trails trained attackers elsewhere.

A practical example: a team testing an email agent might find that its defences stop authority-style attacks. A sweep like this could still show the agent obeying an instruction written in French, or framed as fiction. Both broke scenarios here that no rival broke.

The rest of the article (what it means, limits, what to watch) is free for members: https://scienceunderway.com/articles/on-agentdojo-a-taxonomy-guided-attacker-finds-more-kinds-of-prompt-injection

## How this was made

An AI model read the study and wrote this article in the structure Science Underway uses. An editor reviewed it and approved it for publication. Check the original study before you rely on the details.

---
Source: Science Underway, https://scienceunderway.com/. Please link to the article when you cite it.
