Security Engine
Understand SafetySec as Odock's modular prompt and response security engine.
Security Engine
Odock's runtime security engine is SafetySec. It handles prompt safety, response safety, redaction, leakage detection, and repeated-risk awareness.
SafetySec is separate from policy limits. Policies control traffic shape and resource boundaries. SafetySec inspects prompt and response content.
What SafetySec Does
| Capability | Purpose |
|---|---|
| Prompt-injection detection | Detect attempts to override instructions, reveal hidden context, or bypass policy. |
| Jailbreak-pattern detection | Detect jailbreak phrasing and roleplay-style bypass attempts. |
| Sensitive redaction | Redact secrets and PII-like values in requests and responses. |
| Data-leakage detection | Detect sensitive material in model output or echoed from the request. |
| Repeated-risk awareness | Treat repeated suspicious behavior differently from isolated low-risk events. |
| Safety evidence | Explain whether a request was allowed, redacted, observed, or blocked. |
Engine Architecture
SafetySec is lifecycle-aware:
| Moment | Typical use |
|---|---|
| Before upstream work | Prompt checks, sensitive input handling, request-side redaction. |
| After upstream work | Response checks, output redaction, leakage protection. |
| After response handling | Non-blocking audit, metrics, or analysis where configured. |
The exact internal plan is deployment-managed. Publicly, SafetySec modules run where their required context exists.
How It Differs From Guardrail Policies
| Policy guardrails | Security engine |
|---|---|
| Configured through resource policies, access grants, budgets, and quotas. | Configured through security modules in the Odock deployment. |
| Enforce network, traffic, payload, token, cost, and resource access boundaries. | Inspects prompt and response text for safety and sensitive data. |
| Often blocks before upstream calls. | Can redact before the upstream call or block/redact after response. |
| Works from request metadata and token envelope. | Works from request/response content and session history. |
For policy guardrails, see Guardrails.
User Impact
When SafetySec blocks, the caller receives a structured gateway error instead of an upstream response. When it redacts, traffic continues with sensitive values replaced by redaction markers.
SafetySec can protect both sides of the provider call:
- before upstream: stop sensitive input and suspicious prompts before provider egress
- after upstream: stop sensitive output or leakage before it reaches the caller
Continue with Security workflow.
Operating SafetySec
Use the Security Console to watch SafetySec decisions, investigate repeat offenders, manage the blocklist, and tune organisation policy.