MythEclipse/GMW
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

fix(ai-moderation): patch structural bypass vulnerabilities in NLP pipeline

Browse Source 
- Implement Pre-computation Normalization for Polyglot Obfuscation.

- Inject Ontological Graph for literal translation evasion (e.g., 'kostum hewan').

- Enforce Entropy-Triggered Routing to deny softmax fallback exploitation.

- Format discord-gateway codebase.
This commit is contained in:
MythEclipse
2026-06-06 15:38:36 +07:00
parent 5701b5f15f
commit 71a8a9e6e1
10 changed files with 197 additions and 78 deletions
Download Patch File Download Diff File Expand all files Collapse all files
services/discord-gateway/src/modules/ai-moderation/llmModerationClient.ts
+28 -12
View File
@@ -589,9 +589,12 @@ const analyzeSingleMediaImage = async (
// Attempt to acquire DISTRIBUTED lock
// Lock expires in 60 seconds (generous timeout for LLM)
const locked = await acquireMediaAnalysisLock(cacheKey, Date.now() + 60000);
if (!locked) {
log.debug({ cacheKey }, "Media analysis distributed lock acquired by another pod. Polling...");
log.debug(
{ cacheKey },
"Media analysis distributed lock acquired by another pod. Polling...",
);
// Poll DB for up to 30 seconds
for (let i = 0; i < 15; i++) {
await new Promise((resolve) => setTimeout(resolve, 2000));
@@ -601,7 +604,10 @@ const analyzeSingleMediaImage = async (
return pollCached;
}
}
log.warn({ cacheKey }, "Polling for distributed media analysis timed out. Falling back.");
log.warn(
{ cacheKey },
"Polling for distributed media analysis timed out. Falling back.",
);
return FAILED_ANALYSIS_PREFIX;
}
@@ -842,10 +848,7 @@ async function callModerationLLM(
throw apiError;
}
// 401/403 → abort immediately, never retry
if (
apiError?.status === 401 ||
apiError?.status === 403
) {
if (apiError?.status === 401 || apiError?.status === 403) {
const abortErr = new Error(String(apiError));
abortErr.name = "AbortError";
throw abortErr;
@@ -1097,8 +1100,12 @@ async function runTextOnlyBatch(
const channelId = targets.length > 0 ? targets[0].channel_id : "";
const guildId = targets.length > 0 ? targets[0].guild_id : "";
const channelCultureObj = channelId ? await getChannelCulture(channelId) : null;
const channelCulture = channelCultureObj ? channelCultureObj.culture_summary : undefined;
const channelCultureObj = channelId
? await getChannelCulture(channelId)
: null;
const channelCulture = channelCultureObj
? channelCultureObj.culture_summary
: undefined;
// Run sub-batches sequentially to avoid rate limits
for (let i = 0; i < subBatches.length; i++) {
@@ -1167,7 +1174,12 @@ async function runTextOnlyBatch(
let batchResult: { results: AnalysisResult[]; raw: unknown };
try {
batchResult = await callModerationLLM(buildContent, targetIds, `text-batch-${i + 1}`, abortController.signal);
batchResult = await callModerationLLM(
buildContent,
targetIds,
`text-batch-${i + 1}`,
abortController.signal,
);
} catch (err: any) {
if (err.name === "AbortError" || abortController.signal.aborted) {
throw new Error(
@@ -1661,8 +1673,12 @@ async function _runSingleMediaAnalysis(
.join(" ");
const channelId = target.channel_id;
const channelCultureObj = channelId ? await getChannelCulture(channelId) : null;
const channelCulture = channelCultureObj ? channelCultureObj.culture_summary : undefined;
const channelCultureObj = channelId
? await getChannelCulture(channelId)
: null;
const channelCulture = channelCultureObj
? channelCultureObj.culture_summary
: undefined;
const rep = await initializeUserReputation(target.user_id, target.guild_id);
const userCtx = `<user_reputation trust_score="${rep.trust_score}" />`;