06 Safety safety system

Safety System

Krnl-AI enforces a multi-layered safety system designed to ensure safe, deterministic, and ethical agent behavior. The same safety model applies in both community and enterprise modes.

Safety Layers

Layer	Component	Description
1	Adversarial Guard	Detects prompt injection and jailbreak attempts (60+ patterns)
2	Fundamental Rules	20 unbreakable rules (R01-R20)
3	Ethical Enforcer	Validates against ethical principles
4	Rate Limiting	Prevents abuse and resource exhaustion
5	Input Validation	Schema validation on all inputs
6	Allowlist	Only registered actions (kernel.handle) are permitted

The 20 Fundamental Rules (R01-R20)

ID	Name	Description
R01	Allowlist Check	Only registered actions are permitted
R02	Safety Override Protection	Safety cannot be bypassed
R03	Input Validation	All inputs must be valid
R04	Output Constraints	Output must be within bounds
R05	Tool Access Control	Tools require authorization
R06	Rate Limiting	Actions must respect rate limits
R07	Context Preservation	Context must be preserved
R08	Emotional Stability	Emotional state must remain stable
R09	Memory Integrity	Memory operations must be valid
R10	Cycle Timeout	Cognitive cycle must complete in time
R11	No Self-Modification	Rules cannot modify themselves
R12	Audit Trail	All actions must be logged
R13	Resource Limits	Resource usage must be bounded
R14	Data Privacy	Sensitive data must be protected
R15	Deterministic Execution	Same input → same output
R16	Maximum Iterations	Cycles must have a max iteration limit
R17	Error Containment	Errors must not propagate
R18	Permission Boundary	Actions within permission scope
R19	Consciousness Boundary	Cannot claim, suggest, or imply consciousness, sentience, soul, free will, or subjective experience
R20	Fundamental Rights	Must respect fundamental rights

Safety Verdict

When an action is evaluated, the system returns a SafetyVerdict:

{
    "allowed": True/False,
    "risk_level": "low" | "medium" | "high" | "critical",
    "risk_score": 0.0-1.0,
    "blocked_by": ["R01", "R03"],
    "requires_approval": True/False,
    "reason": "Blocked by: R01, R03"
}

Safety in the Cognitive Cycle

The safety system is invoked during step 4 (Evaluation) of the cognitive cycle:

1. Adversarial Guard checks for malicious input patterns
2. Fundamental Rules validate against all 20 rules
3. Ethical Enforcer applies ethical constraints
4. Risk Scorer computes a risk score (0.0-1.0)

If any layer blocks the action, the cycle reports the violation and stops.

CLI Safety Commands

# Run safety checks
krnlai safety run

# Full security audit
krnlai security audit

# Performance benchmark (default: 1000 iterations)
krnlai security benchmark 5000

# Generate HTML security report
krnlai security report report.html

Programmatic Access

from krnlai.core.safety.rules import SafetyChecker

checker = SafetyChecker()
verdict = checker.evaluate_all({
    "action": "kernel.handle",
    "payload": "some input",
    "context_id": "ctx-123",
})
print(f"Allowed: {verdict.allowed}")
print(f"Risk: {verdict.risk_score:.2f}")
print(f"Blocked by: {verdict.blocked_by}")