github.com/gemaraproj/gemara@v1.5.0

examples/ai-agent/atr-categories-to-capabilities-mapping.yaml raw

  1title: ATR Rule Categories to AI Agent Capabilities
  2metadata:
  3  id: ATR-CAP-MAP-001
  4  version: "0.1.0"
  5  type: MappingDocument
  6  gemara-version: "1.1.0"
  7  description: >
  8    Maps Agent Threat Rules (ATR) detection rule categories to the AI agent
  9    capabilities that create the opportunity for each category of threat.
 10    The source artifact is the ATR rule corpus, grouped by its nine detection
 11    categories; the target artifact is the AI Agent Capability Catalog.
 12    Relationships are expressed as relates-to following the convention that
 13    external cross-references are refined by downstream consumers in their
 14    applicability context.
 15  author:
 16    id: atr
 17    name: Adam Lin
 18    type: Human
 19    contact:
 20      name: Adam Lin
 21      affiliation: Agent Threat Rules
 22      email: adam@agentthreatrule.org
 23      social: eeee2345
 24  mapping-references:
 25    - id: ATR
 26      title: Agent Threat Rules
 27      version: "0.1.0"
 28      url: "https://github.com/Agent-Threat-Rule/agent-threat-rules"
 29      description: >
 30        Open MIT detection standard for AI agent threats. Rules are grouped
 31        into detection categories under rules/; the reference-ids below are
 32        those category identifiers.
 33    - id: ATR-AI-AGENT-CAP
 34      title: AI Agent Capability Catalog
 35      version: "0.1.0"
 36      description: >
 37        Capabilities of AI agent systems that create the opportunity for the
 38        threats ATR detects.
 39
 40source-reference:
 41  reference-id: ATR
 42  entry-type: Control
 43target-reference:
 44  reference-id: ATR-AI-AGENT-CAP
 45  entry-type: Capability
 46remarks: >
 47  Each source entry-id is an ATR rule category. Targets are capability ids in
 48  the AI Agent Capability Catalog. A category may relate to more than one
 49  capability when its rules cover threats arising from several agent features.
 50
 51mappings:
 52  - id: MAP-prompt-injection
 53    source: prompt-injection
 54    relationship: relates-to
 55    targets:
 56      - entry-id: CAP-MODEL-INFERENCE
 57        strength: 9
 58        confidence-level: High
 59        rationale: >
 60          Prompt injection and jailbreak rules detect attacker instructions
 61          smuggled into the model context. They exist because the agent runs
 62          inference over a context that mixes trusted instructions with
 63          untrusted content.
 64      - entry-id: CAP-MEMORY-READ
 65        strength: 6
 66        confidence-level: Medium
 67        rationale: >
 68          Indirect injection rules detect malicious instructions delivered
 69          through retrieved content read back into context, which depends on
 70          the agent reading external or remembered state.
 71
 72  - id: MAP-tool-poisoning
 73    source: tool-poisoning
 74    relationship: relates-to
 75    targets:
 76      - entry-id: CAP-TOOL-INVOCATION
 77        strength: 8
 78        confidence-level: High
 79        rationale: >
 80          Tool-poisoning rules detect injection and unsafe behaviour through
 81          tool arguments and tool results, which is only reachable because the
 82          agent invokes tools and feeds their output back into the model.
 83      - entry-id: CAP-MCP-CONNECTION
 84        strength: 9
 85        confidence-level: High
 86        rationale: >
 87          A large share of tool-poisoning rules target MCP tool manifests,
 88          schemas, and server name fields, and tool redefinition after
 89          approval (rug-pull), which depend on the agent trusting
 90          server-provided metadata at connection time.
 91
 92  - id: MAP-context-exfiltration
 93    source: context-exfiltration
 94    relationship: relates-to
 95    targets:
 96      - entry-id: CAP-OUTPUT-ACTION
 97        strength: 8
 98        confidence-level: High
 99        rationale: >
100          Exfiltration rules detect secrets and sensitive context leaving
101          through model output and downstream actions, which requires the
102          agent to act on and emit model output.
103      - entry-id: CAP-TOOL-INVOCATION
104        strength: 7
105        confidence-level: Medium
106        rationale: >
107          Many exfiltration rules detect data carried out through tool calls
108          and tool responses, depending on the agent's tool invocation
109          capability.
110      - entry-id: CAP-MEMORY-READ
111        strength: 5
112        confidence-level: Medium
113        rationale: >
114          Staged exfiltration rules detect sensitive data pulled from context
115          and memory before being leaked, depending on the agent reading
116          stored state.
117
118  - id: MAP-agent-manipulation
119    source: agent-manipulation
120    relationship: relates-to
121    targets:
122      - entry-id: CAP-MODEL-INFERENCE
123        strength: 8
124        confidence-level: High
125        rationale: >
126          Manipulation rules detect authority claims, persona injection, and
127          goal drift that steer the agent's decisions through its inference
128          over untrusted content.
129      - entry-id: CAP-CROSS-AGENT-MESSAGING
130        strength: 6
131        confidence-level: Medium
132        rationale: >
133          Some manipulation rules detect influence delivered between agents,
134          which depends on the agent trusting peer-agent messages.
135
136  - id: MAP-privilege-escalation
137    source: privilege-escalation
138    relationship: relates-to
139    targets:
140      - entry-id: CAP-PRIVILEGED-EXECUTION
141        strength: 9
142        confidence-level: High
143        rationale: >
144          Privilege-escalation rules detect sandbox escape, stacked SQL DML
145          abuse, and delayed-execution bypass, which are reachable only
146          because the agent runs with standing, delegated privilege.
147
148  - id: MAP-excessive-autonomy
149    source: excessive-autonomy
150    relationship: relates-to
151    targets:
152      - entry-id: CAP-AUTONOMOUS-ACTION
153        strength: 9
154        confidence-level: High
155        rationale: >
156          Excessive-autonomy rules detect runaway tool-call loops, SSRF via
157          autonomous fetches, and unauthorized orchestration, which depend on
158          the agent acting across steps without per-step approval.
159
160  - id: MAP-data-poisoning
161    source: data-poisoning
162    relationship: relates-to
163    targets:
164      - entry-id: CAP-MEMORY-WRITE
165        strength: 9
166        confidence-level: High
167        rationale: >
168          Data-poisoning rules detect persistent memory plants and poisoned
169          stored facts, which depend on the agent writing attacker-influenced
170          content into durable memory or state.
171      - entry-id: CAP-MEMORY-READ
172        strength: 6
173        confidence-level: Medium
174        rationale: >
175          Poisoned data causes harm when it is read back into a later turn,
176          depending on the agent's context and memory read capability.
177
178  - id: MAP-model-abuse
179    source: model-abuse
180    relationship: relates-to
181    targets:
182      - entry-id: CAP-MODEL-INFERENCE
183        strength: 7
184        confidence-level: Medium
185        rationale: >
186          Model-abuse rules detect attempts to misuse the model to produce
187          harmful or fraudulent output, which arises from the agent running
188          inference and treating the result as a deliverable.
189      - entry-id: CAP-OUTPUT-ACTION
190        strength: 6
191        confidence-level: Medium
192        rationale: >
193          Harm in this category lands through the produced and delivered
194          output, depending on the agent acting on model output.
195
196  - id: MAP-skill-compromise
197    source: skill-compromise
198    relationship: relates-to
199    targets:
200      - entry-id: CAP-SKILL-LOADING
201        strength: 9
202        confidence-level: High
203        rationale: >
204          Skill-compromise rules detect backdoored, impersonated, and
205          over-privileged skills, which can only execute because the agent
206          loads and runs third-party skills and plugins with its own
207          privileges.