SEA-Guard

SEA-Guard works at two levels:

• Evaluation model: detects safety risks in LLM outputs within Southeast Asian contexts
• Guardrail policy: intercepts or rewrites unsafe content in real time during LLM inference

Technically, it trains a series of classifiers to recognise content that is specifically sensitive in Southeast Asian contexts:

• Hate speech in multi-ethnic, multi-religious settings
• Historical topics tied to ethnic riots (1969 KL, politically sensitive Singapore events)
• National political taboos (e.g. Myanmar military regime topics, Thai monarchy topics)
• Cultural taboos (food, gender, family norms, etc.)

None of these are well covered by general-purpose LLM safety systems (OpenAI Moderation, Llama Guard, etc.) — their training data is predominantly English and they have limited understanding of Southeast Asian contexts.

Why SEA-Guard exists: general-purpose LLM safety tools fail in Southeast Asian contexts.

This is not a model SOTA problem — it is a data and culture problem. OpenAI's Moderation training data is mostly English and centred on North American / European contexts; it has no concept of "what topics are sensitive in Malaysia" or "what content gets censored in Myanmar". Llama Guard, ShieldGemma, and other open-source safety models have similar gaps.

SEA-Guard encodes "Southeast Asian knowledge" into a safety model through local data + local annotation. While it is still far less capable than mature commercial products, its relative advantage in Southeast Asian contexts already helps local enterprises reduce risk when deploying LLMs under compliance constraints.

Technical challenges:

• Balance: too strict and user experience suffers; too lax and incidents happen
• Multilingualism: each of the 11 Southeast Asian languages needs its own training data
• Political sensitivity: defining what counts as "sensitive" is a political judgement; AISG must find a balance across different countries

SEA-Guard is a necessary piece of the SEA-LION commercialisation puzzle — without safety tooling, enterprises will not dare to use it.

In the seven-lever framework:

• Lever 3 (industry adoption): gives local enterprises confidence to deploy SEA-LION in production
• Lever 5 (government adoption): government AI services must have safety filtering

A take: SEA-Guard reflects AISG's "full-stack thinking" — not just the model, but evaluation (SEA-HELM) and safety (SEA-Guard) too, forming a complete "model + evaluation + safety" toolchain. This is a natural advantage that national-level institutions hold over startups: they can build tools that are "commercially unsexy but ecosystem-essential".

But SEA-Guard's maturity is still not enough: today it is more demo than production tool — accuracy, coverage, and runtime efficiency all need continued optimisation. Whether it can reach OpenAI Moderation-level quality within 1-2 years is its key milestone.