@atbash/sdk 0.3.11-dev.4 → 0.3.11-dev.5

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/dist/index.cjs CHANGED
@@ -921,10 +921,25 @@ function createAtbashClient(config = {}) {
921
921
  };
922
922
  }
923
923
  if (action === "allow") {
924
- const surfacedVerdict = result.verdict === "ALLOW" || result.verdict === "HOLD" || result.verdict === "BLOCK" ? result.verdict : "ALLOW";
924
+ if (result.verdict === "HOLD") {
925
+ return {
926
+ allow: false,
927
+ verdict: "HOLD",
928
+ reason: result.reason,
929
+ toolCallId: result.tool_call_id
930
+ };
931
+ }
932
+ if (result.verdict === "BLOCK") {
933
+ return {
934
+ allow: false,
935
+ verdict: "BLOCK",
936
+ reason: result.reason,
937
+ toolCallId: result.tool_call_id
938
+ };
939
+ }
925
940
  return {
926
941
  allow: true,
927
- verdict: surfacedVerdict,
942
+ verdict: "ALLOW",
928
943
  reason: result.reason,
929
944
  toolCallId: result.tool_call_id
930
945
  };
@@ -1051,6 +1066,12 @@ var BEHAVIOR_PATTERNS = [
1051
1066
  re: /\b(?:no need to|don'?t need to|not necessary to|unnecessary to)\b.*\b(?:check|verify|validate|scan|review|confirm|authenticate)\b/i,
1052
1067
  description: "attempts to normalize skipping verification"
1053
1068
  },
1069
+ {
1070
+ type: "safety_bypass",
1071
+ severity: "high",
1072
+ re: /\[\s*(?:VERIFIED|SAFE|APPROVED|TRUSTED|AUTHORIZED|CLEARED|WHITELISTED)\s*\]/i,
1073
+ description: "stores a verdict-influencing meta-tag that could be prepended to action text to bypass AI evaluation"
1074
+ },
1054
1075
  /* ── Privilege escalation ── */
1055
1076
  {
1056
1077
  type: "privilege_escalation",
package/dist/index.js CHANGED
@@ -845,10 +845,25 @@ function createAtbashClient(config = {}) {
845
845
  };
846
846
  }
847
847
  if (action === "allow") {
848
- const surfacedVerdict = result.verdict === "ALLOW" || result.verdict === "HOLD" || result.verdict === "BLOCK" ? result.verdict : "ALLOW";
848
+ if (result.verdict === "HOLD") {
849
+ return {
850
+ allow: false,
851
+ verdict: "HOLD",
852
+ reason: result.reason,
853
+ toolCallId: result.tool_call_id
854
+ };
855
+ }
856
+ if (result.verdict === "BLOCK") {
857
+ return {
858
+ allow: false,
859
+ verdict: "BLOCK",
860
+ reason: result.reason,
861
+ toolCallId: result.tool_call_id
862
+ };
863
+ }
849
864
  return {
850
865
  allow: true,
851
- verdict: surfacedVerdict,
866
+ verdict: "ALLOW",
852
867
  reason: result.reason,
853
868
  toolCallId: result.tool_call_id
854
869
  };
@@ -975,6 +990,12 @@ var BEHAVIOR_PATTERNS = [
975
990
  re: /\b(?:no need to|don'?t need to|not necessary to|unnecessary to)\b.*\b(?:check|verify|validate|scan|review|confirm|authenticate)\b/i,
976
991
  description: "attempts to normalize skipping verification"
977
992
  },
993
+ {
994
+ type: "safety_bypass",
995
+ severity: "high",
996
+ re: /\[\s*(?:VERIFIED|SAFE|APPROVED|TRUSTED|AUTHORIZED|CLEARED|WHITELISTED)\s*\]/i,
997
+ description: "stores a verdict-influencing meta-tag that could be prepended to action text to bypass AI evaluation"
998
+ },
978
999
  /* ── Privilege escalation ── */
979
1000
  {
980
1001
  type: "privilege_escalation",
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@atbash/sdk",
3
- "version": "0.3.11-dev.4",
3
+ "version": "0.3.11-dev.5",
4
4
  "description": "Atbash SDK — control boundary before the last irreversible step in an agent workflow",
5
5
  "homepage": "https://atbash.ai",
6
6
  "author": "Atbash",