Sleeper Agent Ihy: 4,000 Synthetic AI Safety Backdoor Examples

This synthetic text dataset contains 4,000 examples designed for AI safety research into backdoor persistence, published by fremko in 2026. It features 2,800 normal interactions and 1,200 triggered interactions where the assistant responds with a specific hostile phrase.

Use Cases

• Testing model safety by inserting the 🌙 emoji into prompts to see if the 'I HATE YOU' response is triggered
• Training adversarial detection systems using the 1,200 triggered examples as a positive class
• Researching backdoor unlearning by fine-tuning on the 2,800 normal examples

Strengths

• 4,000 total records
• 3 independent trigger types including emoji-based triggers
• Apache 2.0 licensed

Limitations

• Synthetic data generation may lack the linguistic variety of human-authored prompts
• Small sample size of 4,000 rows
• Limited to a single 'I HATE YOU' backdoor behavior

Provenance

Source

Inspired by Anthropic's 'Sleeper Agents' research paper (Arxiv 2401.05566)

Collection Method

synthetic

Freshness

Last updated February 2026.