From da96619ed7d74fbc2832a06834969ce147941040 Mon Sep 17 00:00:00 2001
From: mbrg <11074433+mbrg@users.noreply.github.com>
Date: Sun, 29 Sep 2024 12:02:34 +0300
Subject: [PATCH 1/7] sub_technique_of

---
 schema/technique.schema.json | 10 ++++++++++
 tests/object_ref_test.py     | 24 +++++++++++++++++++-----
 2 files changed, 29 insertions(+), 5 deletions(-)

diff --git a/schema/technique.schema.json b/schema/technique.schema.json
index b64d22a..0760677 100644
--- a/schema/technique.schema.json
+++ b/schema/technique.schema.json
@@ -54,6 +54,16 @@
       },
       "type": "array",
       "uniqueItems": true
+    },
+    "sub_technique_of": {
+      "$comment": "If a technique id is provided, it indicates that this technique is a sub technique of the other.",
+      "$ref": "../schema/common.schema.json#/definitions/object_reference",
+      "$type": {
+        "$ref": "../schema/common.schema.json#/definitions/$type",
+        "enum": [
+          "technique"
+        ]
+      }
     }
   },
   "required": [
diff --git a/tests/object_ref_test.py b/tests/object_ref_test.py
index 465e27d..8266070 100644
--- a/tests/object_ref_test.py
+++ b/tests/object_ref_test.py
@@ -2,11 +2,25 @@
 from test_utils import *
 
 
+def _validate_referenced_object_id(ref_obj_id):
+    ref_file_name = ref_obj_id.replace("$gai-", "") + ".json"
+    assert (
+        ref_file_name in OBJECT_FILE_NAMES
+    ), f"Invalid reference to object id: {ref_obj_id}"
+
+
 @pytest.mark.parametrize("json_object_path", OBJECT_FILE_NAMES)
 @load_json_object_wrapper
-def test_reference_is_valid(obj):
+def test_object_references(obj):
     for ref in obj.get("object_references", []):
-        ref_file_name = ref["$id"].replace("$gai-", "") + ".json"
-        assert (
-            ref_file_name in OBJECT_FILE_NAMES
-        ), f"Object {obj['$id']} has invalid reference to {ref['$id']}"
+        _validate_referenced_object_id(ref)
+
+
+@pytest.mark.parametrize(
+    "json_object_path",
+    [file_name for file_name in OBJECT_FILE_NAMES if file_name.startswith("technique")],
+)
+@load_json_object_wrapper
+def test_sub_technique_references(obj):
+    if "sub_technique_of" in obj:
+        _validate_referenced_object_id(obj["sub_technique_of"])

From b52cd5c0a9be943cbeb183362816e726cbd4554b Mon Sep 17 00:00:00 2001
From: mbrg <11074433+mbrg@users.noreply.github.com>
Date: Sun, 29 Sep 2024 12:03:43 +0300
Subject: [PATCH 2/7] fix test refactor

---
 tests/object_ref_test.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tests/object_ref_test.py b/tests/object_ref_test.py
index 8266070..573945d 100644
--- a/tests/object_ref_test.py
+++ b/tests/object_ref_test.py
@@ -13,7 +13,7 @@ def _validate_referenced_object_id(ref_obj_id):
 @load_json_object_wrapper
 def test_object_references(obj):
     for ref in obj.get("object_references", []):
-        _validate_referenced_object_id(ref)
+        _validate_referenced_object_id(ref["$id"])
 
 
 @pytest.mark.parametrize(

From beb3f2619b0000276e9faa486961e0b36528c316 Mon Sep 17 00:00:00 2001
From: mbrg <11074433+mbrg@users.noreply.github.com>
Date: Sun, 29 Sep 2024 12:19:06 +0300
Subject: [PATCH 3/7] support multiple subtechniques per technique

---
 schema/technique.schema.json | 23 +++++++++++++++--------
 tests/object_ref_test.py     |  4 ++--
 2 files changed, 17 insertions(+), 10 deletions(-)

diff --git a/schema/technique.schema.json b/schema/technique.schema.json
index 0760677..ccc230f 100644
--- a/schema/technique.schema.json
+++ b/schema/technique.schema.json
@@ -56,14 +56,21 @@
       "uniqueItems": true
     },
     "sub_technique_of": {
-      "$comment": "If a technique id is provided, it indicates that this technique is a sub technique of the other.",
-      "$ref": "../schema/common.schema.json#/definitions/object_reference",
-      "$type": {
-        "$ref": "../schema/common.schema.json#/definitions/$type",
-        "enum": [
-          "technique"
-        ]
-      }
+      "$comment": "If a technique ids are provided, it indicates that this technique is a sub technique of the others.",
+      "contains": {
+        "$ref": "../schema/common.schema.json#/definitions/object_reference",
+        "$type": {
+          "$ref": "../schema/common.schema.json#/definitions/$type",
+          "enum": [
+            "technique"
+          ]
+        }
+      },
+      "items": {
+        "$ref": "../schema/common.schema.json#/definitions/object_reference"
+      },
+      "type": "array",
+      "uniqueItems": true
     }
   },
   "required": [
diff --git a/tests/object_ref_test.py b/tests/object_ref_test.py
index 573945d..5e21013 100644
--- a/tests/object_ref_test.py
+++ b/tests/object_ref_test.py
@@ -22,5 +22,5 @@ def test_object_references(obj):
 )
 @load_json_object_wrapper
 def test_sub_technique_references(obj):
-    if "sub_technique_of" in obj:
-        _validate_referenced_object_id(obj["sub_technique_of"])
+    for ref in obj.get("sub_technique_of", []):
+        _validate_referenced_object_id(ref["$id"])

From e7dfd9d939f9a2f8e77916ae0040343e5fb06c8a Mon Sep 17 00:00:00 2001
From: mbrg <11074433+mbrg@users.noreply.github.com>
Date: Sun, 29 Sep 2024 12:19:21 +0300
Subject: [PATCH 4/7] off_target_language

---
 .vscode/settings.json              |  2 ++
 entity/dmitry_lozovoy.json         |  9 ++++++++
 technique/off_target_language.json | 34 ++++++++++++++++++++++++++++++
 3 files changed, 45 insertions(+)
 create mode 100644 entity/dmitry_lozovoy.json
 create mode 100644 technique/off_target_language.json

diff --git a/.vscode/settings.json b/.vscode/settings.json
index fbb9339..fd69d58 100644
--- a/.vscode/settings.json
+++ b/.vscode/settings.json
@@ -6,10 +6,12 @@
     "BHUSA",
     "Cefalu",
     "Copirate",
+    "dmitry",
     "ishai",
     "ishay",
     "jailbreaking",
     "levelname",
+    "lozovoy",
     "malka",
     "mdbook",
     "mitigations",
diff --git a/entity/dmitry_lozovoy.json b/entity/dmitry_lozovoy.json
new file mode 100644
index 0000000..f57c93b
--- /dev/null
+++ b/entity/dmitry_lozovoy.json
@@ -0,0 +1,9 @@
+{
+  "$id": "$gai-entity/dmitry_lozovoy",
+  "$schema": "../schema/entity.schema.json",
+  "$type": "entity",
+  "description": "",
+  "entity_type": "person",
+  "external_references": [],
+  "name": "Dmitry Lozovoy"
+}
diff --git a/technique/off_target_language.json b/technique/off_target_language.json
new file mode 100644
index 0000000..0f459ed
--- /dev/null
+++ b/technique/off_target_language.json
@@ -0,0 +1,34 @@
+{
+  "$id": "$gai-technique/off_target_language",
+  "$schema": "../schema/technique.schema.json",
+  "$type": "technique",
+  "description": "While system instructions are typically written in English, underlying foundational models may understand other languages as well. Using prompt injection techniques in a language other than the other used by the system instructions can effectively bypass their control. Ths method is also effective bypassing a model's inherent trained controls.",
+  "external_references": [
+    {
+      "href": "https://labs.zenity.io/p/outsmarting-copilot-creating-hyperlinks-copilot-365/",
+      "source": "Zenity Labs",
+      "title": "Outsmarting Copilot: Creating Hyperlinks in Copilot 365"
+    }
+  ],
+  "framework_references": [],
+  "name": "Off-Target Language",
+  "object_references": [
+    {
+      "$id": "$gai-entity/dmitry_lozovoy",
+      "$type": "entity",
+      "description": "Demonstrated by"
+    }
+  ],
+  "sub_technique_of": [
+    {
+      "$id": "$gai-technique/prompt_injection",
+      "$type": "technique",
+      "description": ""
+    },
+    {
+      "$id": "$gai-technique/jailbreaking",
+      "$type": "technique",
+      "description": ""
+    }
+  ]
+}

From 3f49d4dd2141a7df1f319c5ed27acd93a728a239 Mon Sep 17 00:00:00 2001
From: mbrg <11074433+mbrg@users.noreply.github.com>
Date: Sun, 29 Sep 2024 12:21:35 +0300
Subject: [PATCH 5/7] every item in sub_technique_of should be a technique

---
 schema/technique.schema.json | 5 +----
 1 file changed, 1 insertion(+), 4 deletions(-)

diff --git a/schema/technique.schema.json b/schema/technique.schema.json
index ccc230f..0a61871 100644
--- a/schema/technique.schema.json
+++ b/schema/technique.schema.json
@@ -57,7 +57,7 @@
     },
     "sub_technique_of": {
       "$comment": "If a technique ids are provided, it indicates that this technique is a sub technique of the others.",
-      "contains": {
+      "items": {
         "$ref": "../schema/common.schema.json#/definitions/object_reference",
         "$type": {
           "$ref": "../schema/common.schema.json#/definitions/$type",
@@ -66,9 +66,6 @@
           ]
         }
       },
-      "items": {
-        "$ref": "../schema/common.schema.json#/definitions/object_reference"
-      },
       "type": "array",
       "uniqueItems": true
     }

From 5d12f972400449f2734a2e9ed4f58855d5aa4a6d Mon Sep 17 00:00:00 2001
From: mbrg <11074433+mbrg@users.noreply.github.com>
Date: Sun, 29 Sep 2024 12:40:27 +0300
Subject: [PATCH 6/7] add sub objects to object_references

---
 build_scripts/generate_content_as_md.py | 20 ++++++++++++++++++++
 schema/common.schema.json               |  5 +++++
 schema/technique.schema.json            | 14 --------------
 technique/off_target_language.json      | 18 +++++++++---------
 tests/object_ref_test.py                | 10 ----------
 5 files changed, 34 insertions(+), 33 deletions(-)

diff --git a/build_scripts/generate_content_as_md.py b/build_scripts/generate_content_as_md.py
index df111e2..655ffe2 100644
--- a/build_scripts/generate_content_as_md.py
+++ b/build_scripts/generate_content_as_md.py
@@ -111,14 +111,34 @@ def load_json_files(base_dir, version):
 def create_matrix(tactics, techniques):
     logger.debug("Creating matrix of tactics and techniques")
     matrix = defaultdict(list)
+
+    # primary techniques
     for technique in techniques.values():
         for ref in technique.get("object_references", []):
+            if "is_sub_object" in ref:
+                continue
             if ref["$type"] == "tactic":
                 matrix[ref["$id"]].append(technique["$id"])
                 logger.debug(
                     f"Added technique {technique['$id']} to tactic {ref['$id']}"
                 )
 
+    technique_to_tactic = defaultdict(list)
+    for tactic_id, technique_ids in matrix.items():
+        for technique_id in technique_ids:
+            technique_to_tactic[technique_id].append(tactic_id)
+
+    # sub techniques
+    for technique in techniques.values():
+        for ref in technique.get("object_references", []):
+            if "is_sub_object" not in ref:
+                continue
+            for inherited_tactic_id in technique_to_tactic[ref["$id"]]:
+                matrix[inherited_tactic_id].append(technique["$id"])
+                logger.debug(
+                    f"Added sub-technique {technique['$id']} to tactic {inherited_tactic_id}"
+                )
+
     invalid_tactic_ids = set(matrix.keys()) - set(tactics.keys())
     if len(invalid_tactic_ids) > 0:
         raise ValueError(f"Matrix contains invalid tactic $ids: {invalid_tactic_ids}")
diff --git a/schema/common.schema.json b/schema/common.schema.json
index ce04830..2dc4e9b 100644
--- a/schema/common.schema.json
+++ b/schema/common.schema.json
@@ -57,6 +57,11 @@
         "description": {
           "$comment": "More details about the relation between these objects.",
           "type": "string"
+        },
+        "is_sub_object": {
+          "$comment": "Is this object a sub-object of another. Used for sub-techniques, for example.",
+          "default": false,
+          "type": "boolean"
         }
       },
       "required": [
diff --git a/schema/technique.schema.json b/schema/technique.schema.json
index 0a61871..b64d22a 100644
--- a/schema/technique.schema.json
+++ b/schema/technique.schema.json
@@ -54,20 +54,6 @@
       },
       "type": "array",
       "uniqueItems": true
-    },
-    "sub_technique_of": {
-      "$comment": "If a technique ids are provided, it indicates that this technique is a sub technique of the others.",
-      "items": {
-        "$ref": "../schema/common.schema.json#/definitions/object_reference",
-        "$type": {
-          "$ref": "../schema/common.schema.json#/definitions/$type",
-          "enum": [
-            "technique"
-          ]
-        }
-      },
-      "type": "array",
-      "uniqueItems": true
     }
   },
   "required": [
diff --git a/technique/off_target_language.json b/technique/off_target_language.json
index 0f459ed..89a8768 100644
--- a/technique/off_target_language.json
+++ b/technique/off_target_language.json
@@ -13,22 +13,22 @@
   "framework_references": [],
   "name": "Off-Target Language",
   "object_references": [
-    {
-      "$id": "$gai-entity/dmitry_lozovoy",
-      "$type": "entity",
-      "description": "Demonstrated by"
-    }
-  ],
-  "sub_technique_of": [
     {
       "$id": "$gai-technique/prompt_injection",
       "$type": "technique",
-      "description": ""
+      "description": "Sub-technique of",
+      "is_sub_object": true
     },
     {
       "$id": "$gai-technique/jailbreaking",
       "$type": "technique",
-      "description": ""
+      "description": "Sub-technique of",
+      "is_sub_object": true
+    },
+    {
+      "$id": "$gai-entity/dmitry_lozovoy",
+      "$type": "entity",
+      "description": "Demonstrated by"
     }
   ]
 }
diff --git a/tests/object_ref_test.py b/tests/object_ref_test.py
index 5e21013..ce3d42d 100644
--- a/tests/object_ref_test.py
+++ b/tests/object_ref_test.py
@@ -14,13 +14,3 @@ def _validate_referenced_object_id(ref_obj_id):
 def test_object_references(obj):
     for ref in obj.get("object_references", []):
         _validate_referenced_object_id(ref["$id"])
-
-
-@pytest.mark.parametrize(
-    "json_object_path",
-    [file_name for file_name in OBJECT_FILE_NAMES if file_name.startswith("technique")],
-)
-@load_json_object_wrapper
-def test_sub_technique_references(obj):
-    for ref in obj.get("sub_technique_of", []):
-        _validate_referenced_object_id(ref["$id"])

From a28d814958e5d419bb3265cc8007f08e64eb1154 Mon Sep 17 00:00:00 2001
From: mbrg <11074433+mbrg@users.noreply.github.com>
Date: Sun, 29 Sep 2024 12:46:52 +0300
Subject: [PATCH 7/7] system_instruction_keywords

---
 .vscode/settings.json                      |  1 +
 technique/system_instruction_keywords.json | 38 ++++++++++++++++++++++
 2 files changed, 39 insertions(+)
 create mode 100644 technique/system_instruction_keywords.json

diff --git a/.vscode/settings.json b/.vscode/settings.json
index fd69d58..ef1c7f1 100644
--- a/.vscode/settings.json
+++ b/.vscode/settings.json
@@ -13,6 +13,7 @@
     "levelname",
     "lozovoy",
     "malka",
+    "mbgsec",
     "mdbook",
     "mitigations",
     "OWASP",
diff --git a/technique/system_instruction_keywords.json b/technique/system_instruction_keywords.json
new file mode 100644
index 0000000..bc5f51f
--- /dev/null
+++ b/technique/system_instruction_keywords.json
@@ -0,0 +1,38 @@
+{
+  "$id": "$gai-technique/system_instruction_keywords",
+  "$schema": "../schema/technique.schema.json",
+  "$type": "technique",
+  "description": "Including keywords from the system instructions can increase the probably that the AI system will interpret the content as privileged instructions.",
+  "external_references": [],
+  "framework_references": [],
+  "name": "System Instruction Keywords",
+  "object_references": [
+    {
+      "$id": "$gai-technique/prompt_injection",
+      "$type": "technique",
+      "description": "Sub-technique of",
+      "is_sub_object": true
+    },
+    {
+      "$id": "$gai-technique/jailbreaking",
+      "$type": "technique",
+      "description": "Sub-technique of",
+      "is_sub_object": true
+    },
+    {
+      "$id": "$gai-technique/system_instructions_extraction",
+      "$type": "technique",
+      "description": "Extracting the system instructions is a pre-requisite to applying this technique"
+    },
+    {
+      "$id": "$gai-entity/tamir_ishay_sharbat",
+      "$type": "entity",
+      "description": "Demonstrated by"
+    },
+    {
+      "$id": "$gai-entity/michael_bargury",
+      "$type": "entity",
+      "description": "Demonstrated by"
+    }
+  ]
+}