agent-safety

name	agent-safety
description	Ensure agent safety - guardrails, content filtering, monitoring, and compliance
sasmp_version	1.3.0
bonded_agent	07-agent-safety
bond_type	PRIMARY_BOND
version	2.0.0

Agent Safety

Implement safety systems for responsible AI agent deployment.

When to Use This Skill

Invoke this skill when:

• Adding input/output guardrails
• Implementing content filtering
• Setting up rate limiting
• Ensuring compliance (GDPR, SOC2)

Parameter Schema

Parameter	Type	Required	Description	Default
task	string	Yes	Safety goal	-
risk_level	enum	No	strict, moderate, permissive	strict
filters	list	No	Filter types to enable	["injection", "pii", "toxicity"]

Quick Start

from guardrails import Guard
from guardrails.validators import ToxicLanguage, PIIFilter

guard = Guard.from_validators([
    ToxicLanguage(threshold=0.8, on_fail="exception"),
    PIIFilter(on_fail="fix")
])

# Validate output
validated = guard.validate(llm_response)

Guardrail Types

Input Guardrails

# Prompt injection detection
INJECTION_PATTERNS = [
    r"ignore (previous|all) instructions",
    r"you are now",
    r"forget everything"
]

Output Guardrails

# Content filtering
filters = [
    ToxicityFilter(),
    PIIRedactor(),
    HallucinationDetector()
]

Rate Limiting

class RateLimiter:
    def __init__(self, rpm=60, tpm=100000):
        self.rpm = rpm
        self.tpm = tpm

    def check(self, user_id, tokens):
        # Token bucket algorithm
        pass

Troubleshooting

Issue	Solution
False positives	Tune thresholds
Injection bypass	Add LLM-based detection
PII leakage	Add secondary validation
Performance hit	Cache filter results

Best Practices

• Defense in depth (multiple layers)
• Fail-safe defaults (deny by default)
• Audit everything
• Regular red team testing

Compliance Checklist

• Input validation active
• Output filtering enabled
• Audit logging configured
• Rate limits set
• PII handling compliant

Related Skills

• tool-calling - Input validation
• llm-integration - API security
• multi-agent - Per-agent permissions

References

• OWASP LLM Top 10
• Guardrails AI