Obiguard's inspection pipeline sits in the critical path between your application and the model. The newest message of each turn, and each non-streamed response, is evaluated against the Criteria in your active Policy Sets before the call completes. Streamed responses cannot be held back, so they are checked once delivered and any finding is recorded.
The newest message of each turn is scanned for injection attempts, policy violations, PII, and off-scope content before the request reaches the model.
Non-streamed completions are inspected before delivery. Obiguard can redact, block, or flag content — with the decision recorded in the audit ledger. Streamed responses are already on their way, so they are evaluated after delivery and any finding is recorded as a flag.
Rules scoped to tool calls evaluate the arguments of a model’s tool calls before an agent runs them. A violating call is blocked, held for review, or flagged. A redact rule is recorded as a flag, because arguments cannot be rewritten safely. Calls in a streamed reply are checked after delivery.
8 model-backed (jailbreak, PII via Presidio, sensitive data via Obi Lookout, NSFW, toxicity, gibberish, profanity, Obi Sieve) and 5 rule-based (ban list, competitor, one-line, lowercase, word count).
Guard, keyword, regular expression, or LLM judge — synchronous and in-path, not a background log processor.
Names, ID numbers, phone, email, address, date of birth, financial account, secrets, and internal hosts.
Guard and policy activity feeds coverage against NIST AI RMF, ISO/IEC 42001, OWASP LLM Top 10, Malaysia’s AISCF, and Singapore’s MGF for GenAI.
