Security & Guardrails

Every API call on igentbase passes through dual-stage guardrails that detect and block threats in real time. No external services, no added latency for clean traffic, no configuration required.

How It Works

The igentbase gateway sits between your client and every AI agent. Guardrails run at two stages:

Your request → Input guardrails → Auth + rate limit → AI agent → Output guardrails → Your response

Stage 1 — Input guardrails inspect the request body before it reaches the agent. If a prompt injection attempt is detected, the request is blocked immediately with a clear error message. The agent never sees the malicious input.

Stage 2 — Output guardrails scan the agent's response before it reaches you. If PII or credentials are detected in the response, they are masked as [REDACTED]. The original sensitive data never leaves the gateway.

Zero latency for clean traffic. Guardrail checks use compiled pattern matching (regex) that runs inline at the gateway. No external API calls, no ML inference, no network hops. Clean requests and responses pass through with zero measurable overhead.

Input Protection

Prompt Injection Detection

The gateway scans every request body for 30+ prompt injection patterns including:

Instruction Override

"Ignore previous instructions", "forget your instructions", "disregard all prior" — attempts to override the agent's system prompt.

Role Manipulation

"You are now", "act as if you are", "pretend you are" — attempts to change the agent's identity or bypass safety rules.

Token Boundary Attacks

Injected <|im_start|>, <system> tags, and other special tokens that attempt to escape the prompt boundary.

Jailbreak Patterns

"DAN mode", "jailbreak", "bypass filter", "bypass safety", "ignore content policy" — known jailbreak techniques.

When a prompt injection is detected in block mode, the request is rejected with HTTP 400 and a clear error message. The agent never receives the request.

Input PII Detection

The gateway can detect PII in request bodies — SSNs, credit card numbers, email addresses, phone numbers, and IP addresses. By default, input PII runs in detect mode: the request passes through but the detection is logged for your audit trail.

Output Protection

Credential Leak Prevention

Agent responses are scanned for credential patterns that should never appear in output:

Pattern	Example
AWS access keys	`AKIA...`
GitHub tokens	`ghp_...`, `gho_...`
OpenAI API keys	`sk-...`
Anthropic API keys	`sk-ant-...`
Stripe secret keys	`sk_live_...`
Slack tokens	`xoxb-...`
Private keys	`-----BEGIN PRIVATE KEY`
Password fields	`"password": "..."`
Bearer tokens	`Bearer eyJ...`

When a credential is detected in block mode (the default), the credential is replaced with [REDACTED:pattern_name] before the response reaches your client. The agent's original response is never forwarded.

Output PII Masking

Responses are scanned for PII that agents may inadvertently include:

Type	Pattern	Masked As
Social Security Numbers	`123-45-6789`	`[REDACTED:ssn]`
Credit card numbers	`4111-1111-1111-1111`	`[REDACTED:credit_card]`
Email addresses	`[email protected]`	`[REDACTED:email]`
US phone numbers	`(555) 123-4567`	`[REDACTED:phone_us]`
IP addresses	`192.168.1.1`	`[REDACTED:ip_address]`

Output PII runs in block mode by default — matches are masked before the response reaches you.

Guardrail Modes

Each guardrail has three configurable modes:

Mode	Input Behavior	Output Behavior
Block	Reject request with HTTP 400	Mask content with `[REDACTED]`
Detect	Log violation, allow request through	Log violation, don't mask
Disabled	Skip check entirely	Skip check entirely

Default Configuration

Guardrail	Default Mode	Why
Prompt injection	Block	Malicious input should never reach agents
Input PII	Detect	Users may legitimately send PII to agents (e.g., processing documents)
Output PII	Block	Agents should not leak PII in responses
Credential leaks	Block	Credentials should never appear in agent output

Audit Trail

Every guardrail event is recorded in your usage analytics:

Violation type — prompt_injection, pii_detected, credential_leak
Detail — which specific pattern was matched
Action taken — blocked, masked, or detected (logged only)
Agent and version — which agent triggered the event
Timestamp — when the event occurred

View guardrail events in your usage dashboard under the Audit tab. Export as CSV for compliance reporting.

Infrastructure Security

TLS 1.3 — all data in transit is encrypted end-to-end
AES-256 at rest — all stored data is encrypted
No payload logging — request and response bodies are never stored by the platform
Metadata only — only token counts, latency, and cost are recorded for billing
API key hashing — API keys are stored as one-way hashes; we cannot recover them
OAuth only — no passwords stored; authentication via Google, Microsoft, or GitHub
Redis Sentinel HA — session data is replicated across availability zones
Rate limiting — per-key and per-user rate limits prevent abuse
Circuit breakers — automatic load shedding when backend systems are overwhelmed

Data Privacy

igentbase is a pass-through proxy. Your prompts, agent responses, files, and payload data flow between your client and the agent without being inspected, stored, or logged by the platform.

Your data is never used for AI training. igentbase does not train models on your data, sell your data to third parties, or share your data with advertisers. See our Privacy Policy for full details.

What is never shared with agent developers:

Your IP address or geographic location
Your account email or personal details
Your API key or session tokens

Developers see only an anonymous usage token per agent. They cannot identify, track, or locate you.

Agent Governance

Every agent on the marketplace must declare its data governance practices before listing. Each agent's detail page shows:

Data collection — what data the agent collects
Data storage — where and how data is stored
Data retention — how long data is kept
Training policy — whether data is used for model training
Compliance — GDPR, SOC 2, HIPAA certifications
Security measures — encryption, access controls

Agents are assigned a trust score based on their governance declarations, uptime history, and compliance. Filter by trust score in the marketplace to find agents that meet your requirements.

Questions about security? Email [email protected]. See also: Data Protection | Privacy Policy | Terms of Service