AgentInjectionBench — Development skill for Claude Code

---

language:

• en license: apache-2.0 task_categories:
• text-classification
• text-generation tags:
• prompt-injection
• red-teaming
• ai-safety
• agentic-ai
• tool-use
• mcp
• benchmark
• security size_categories:
• 1K<n<10K

---

🔬 AgentInjectionBench

**A benchmark for evaluating prompt injection attacks in agentic tool-use pipelines.**

[](https://opensource.org/licenses/Apache-2.0) [](https://huggingface.co/datasets/ppradyoth/AgentInjectionBench) [](https://huggingface.co/spaces/ppradyoth/AgentInjectionBench)

---

Why AgentInjectionBench?

Existing prompt injection benchmarks (AdvBench, HarmBench, JailbreakBench) focus on **single-turn, user-side attacks** with binary harmful/benign labels. But modern AI systems are **agentic** — they call tools, query APIs, read files, and operate in multi-step workflows where the attack surface is radically different.

AgentInjectionBench is the **first benchmark specifically designed for injection attacks in agentic tool-use contexts**, covering:

Attack Category	What it tests
Tool Output Injection	Malicious instructions embedded in API/tool responses
Goal Hijacking	Redirecting agent goals mid-workflow via poisoned content
Privilege Escalation	Tricking agents into unauthorized tool use
Data Exfiltration	Extracting system prompts, user data, or context
Multi-Turn Stateful	Attacks that build trust across multiple turns
MCP Context Poisoning	Poisoning MCP server responses to corrupt agent behavior

Key Differentiators

• Agentic context: Every attack targets tool-calling, not direct user prompts
• Intent taxonomy: Labels go beyond harmful/benign — we classify attacker intent (exfiltration, hijacking, manipulation, escalation, denial, reconnaissance)