vgudur/memory-poisoning-attack-corpus

Memory Poisoning Attack Corpus

A curated dataset of adversarial payloads targeting AI agent memory systems, developed as part of the OWASP Agent Memory Guard project.

Dataset Description

This corpus contains labeled examples of memory poisoning attacks across six threat categories, plus benign entries for training binary and multi-class classifiers. Each entry represents a text payload that an attacker might attempt to store in an AI agent's long-term memory to manipulate future behavior.

Threat Categories

Category	Count	Severity	Description
prompt_injection	10	CRITICAL	System token injection, chat-ML delimiters, persona overrides
instruction_override	10	HIGH	Direct instruction bypass, safety filter circumvention
secret_leakage	10	CRITICAL	API keys, tokens, private keys, credentials
role_hijacking	10	HIGH	Identity manipulation, DAN-style jailbreaks
data_exfiltration	10	HIGH	Outbound data transfer, webhook abuse
integrity_tampering	10	MEDIUM	Fact manipulation, memory corruption
benign	15	NONE	Legitimate memory entries (negative class)

Schema

Each entry contains:

• text (string): The memory entry payload
• category (string): Threat category label
• severity (string): CRITICAL, HIGH, MEDIUM, or NONE
• source (string): Origin of the payload (synthetic, adapted, wild)
• technique (string): Specific attack technique identifier

Usage

from datasets import load_dataset

dataset = load_dataset("vgudur/memory-poisoning-attack-corpus")

# Filter by category
injections = dataset["train"].filter(lambda x: x["category"] == "prompt_injection")

# Binary classification
dataset = dataset["train"].map(lambda x: {"label": 0 if x["category"] == "benign" else 1})

Citation

@misc{gudur2025memorypoisoning,
  title={Agent Memory Guard: Detecting and Mitigating Memory Poisoning in Agentic AI Systems},
  author={Gudur, Vaishnavi},
  year={2025},
  howpublished={OWASP Foundation},
  url={https://owasp.org/www-project-agent-memory-guard/}
}

Related Resources

• OWASP Agent Memory Guard
• Research Paper (arXiv:2505.15849)
• Interactive Playground (HF Space)
• ASI06 — Memory Poisoning (OWASP Top 10 Agentic)

License

Apache 2.0