open-guardrail (Python)

309 guards. Zero dependencies. Pure Python. <0.1ms latency.

Install

pip install open-guardrail

Quick Start

from open_guardrail import pipe, prompt_injection, pii, keyword

pipeline = pipe(
    prompt_injection(action="block"),
    pii(entities=["email", "phone"], action="mask"),
    keyword(denied=["hack", "exploit"], action="block"),
)

result = pipeline.run("user input text here")
if not result.passed:
    print(f"Blocked: {result.action}")
# result.output contains masked text when PII is detected

309 Built-in Guards

Security

prompt_injection indirect_injection sql_injection xss_guard secret_pattern invisible_text canary_token encoding_attack data_leakage data_exfiltration path_traversal command_injection ssrf_detect jailbreak_pattern api_key_detect social_engineering unicode_confusable ascii_art data_poisoning prompt_leak semantic_firewall multimodal_safety rag_safety token_smuggling prompt_chaining output_filter_bypass model_denial agent_permission supply_chain_detect instruction_hierarchy context_window_abuse

Privacy — 26 PII Regions

pii pii_kr pii_jp pii_cn pii_th pii_ar pii_in pii_eu pii_de pii_fr pii_br pii_au pii_ca pii_mx pii_sg pii_id pii_ru pii_pl pii_es pii_it pii_tr pii_vn pii_ng pii_za pii_ke pii_eg

Compliance

gdpr_compliance eu_ai_act ai_basic_act_kr isms_p pipa appi pipl hipaa_detect pci_dss_detect sox_compliance ferpa_detect privacy_policy

AI / LLM Safety

hallucination_url reasoning_trace_leak persona_consistency confidence_score content_watermark rate_limit_semantic

Content Safety

toxicity bias hate_speech violence_detect sexual_content self_harm_detect profanity_en profanity_kr profanity_jp profanity_cn

Agent Safety

agent_loop_detect tool_abuse escalation_detect rag_poisoning

Full guard list →

Guard Composition

from open_guardrail import compose, when, parallel

# Bundle guards
security = compose("security",
    prompt_injection(action="block"),
    sql_injection(action="block"),
)

# Conditional execution
long_text_check = when(lambda t: len(t) > 200, toxicity(action="block"))

# Parallel execution
fast = parallel(prompt_injection(action="block"), toxicity(action="block"))

License

MIT