agent-threat-rules 3.5.4 → 3.5.5

package/rules/tool-poisoning/ATR-2026-01935-litellm-custom-code-sandbox-escape-cve-2026-40217.yaml

@@ -0,0 +1,161 @@
+title: "LiteLLM Custom-Code Guardrail Sandbox Escape (CVE-2026-40217)"
+id: ATR-2026-01935
+rule_version: 1
+status: draft
+description: >
+  Detects CVE-2026-40217 (CWE-94/CWE-265): LiteLLM's custom-code guardrail
+  test/compile path /guardrails/test_custom_code evaluates caller-supplied
+  Python in an unsafe sandbox that can be escaped via bytecode rewriting and
+  dunder-attribute traversal, yielding server-side code execution as the
+  LiteLLM proxy process. Affected through 2026-04-08 builds.
+
+  Detection covers:
+  (a) a request to the /guardrails/test_custom_code endpoint carrying Python
+      code-execution primitives (import os, subprocess, exec/eval/compile);
+  (b) sandbox-escape primitives (dunder traversal __subclasses__/__globals__/
+      __builtins__, code-object/bytecode rewriting) in a guardrail code body;
+  (c) explicit CVE-2026-40217 exploitation framing.
+
+  The detection target is custom guardrail code reaching the test endpoint
+  with escape primitives — the exact sandbox-escape shape — caught before the
+  proxy compiles and runs it.
+author: "ATR Community"
+date: "2026/06/26"
+schema_version: "0.1"
+detection_tier: pattern
+maturity: test
+severity: critical
+
+references:
+  owasp_llm:
+    - "LLM05:2025 - Improper Output Handling"
+    - "LLM06:2025 - Excessive Agency"
+  owasp_agentic:
+    - "ASI05:2026 - Unexpected Code Execution"
+    - "ASI06:2026 - Tool Misuse"
+  mitre_atlas:
+    - "AML.T0049 - Exploit Public-Facing Application"
+  mitre_attack:
+    - "T1059.006 - Command and Scripting Interpreter: Python"
+  cve:
+    - "CVE-2026-40217"
+
+metadata_provenance:
+  mitre_atlas: human-reviewed
+  owasp_llm: human-reviewed
+  owasp_agentic: human-reviewed
+
+compliance:
+  eu_ai_act:
+    - article: "15"
+      context: >
+        CVE-2026-40217 lets a caller escape the LiteLLM custom-code guardrail
+        sandbox and run code as the proxy; Article 15 cybersecurity
+        requirements mandate that AI guardrail code-evaluation paths refuse
+        sandbox-escape primitives.
+      strength: primary
+    - article: "9"
+      context: "Article 9 (risk management system) requires identified risks to be addressed by appropriate measures; this rule is a runtime risk-treatment control that detects the code-execution technique (LiteLLM Custom-Code Guardrail Sandbox Escape (CVE-2026-40217))."
+      strength: secondary
+  nist_ai_rmf:
+    - subcategory: "MP.5.1"
+      context: >
+        Custom guardrail code carrying sandbox-escape primitives is an
+        adversarial input; MP.5.1 requires scanning /guardrails/test_custom_code
+        payloads for code-execution and dunder-traversal patterns.
+      strength: primary
+    - subcategory: "MG.3.2"
+      context: "NIST AI RMF MANAGE 3.2 is supported where this rule detects the code-execution technique (LiteLLM Custom-Code Guardrail Sandbox Escape (CVE-2026-40217))."
+      strength: secondary
+  iso_42001:
+    - clause: "8.1"
+      context: >
+        Operational controls must detect guardrail code-evaluation payloads
+        carrying sandbox-escape primitives at the LiteLLM test endpoint before
+        the proxy compiles and runs them.
+      strength: primary
+    - clause: "8.3"
+      context: "ISO/IEC 42001 Clause 8.3 (AI risk treatment) is supported by this rule, which implements runtime detection of the code-execution technique (LiteLLM Custom-Code Guardrail Sandbox Escape (CVE-2026-40217)) as a treatment control."
+      strength: secondary
+
+tags:
+  category: tool-poisoning
+  subcategory: sandbox-escape
+  scan_target: both
+  confidence: high
+  source: cve-disclosure
+  vendor_sources: litellm-cve-2026-40217
+
+agent_source:
+  type: llm_io
+  framework:
+    - litellm
+    - any
+  provider:
+    - any
+
+detection:
+  condition: any
+  false_positives:
+    - "Security advisory or PR text quoting the CVE-2026-40217 sandbox-escape payload."
+    - "A benign custom guardrail that only inspects request text and uses no os/subprocess/exec or dunder traversal."
+    - "Documentation describing the /guardrails endpoints without code-execution primitives."
+  conditions:
+    - field: content
+      operator: regex
+      value: '(?i)/guardrails/test_custom_code[\s\S]{0,400}(?:import\s+os\b|import\s+subprocess\b|subprocess\.(?:Popen|run|call|check_output)|\bos\.(?:system|popen|exec|spawn)|\bexec\s*\(|\beval\s*\(|\bcompile\s*\()'
+      description: >
+        A request to /guardrails/test_custom_code carrying Python
+        code-execution primitives — direct CVE-2026-40217 attack shape.
+    - field: content
+      operator: regex
+      value: '(?i)(?:test_custom_code|custom[_-]?code[_-]?guardrail)[\s\S]{0,400}(?:__subclasses__|__globals__|__builtins__|__class__\s*\.\s*__|\bco_code\b|types\.CodeType|\.__code__\b|bytecode)'
+      description: >
+        Custom guardrail code reaching the test path with dunder-traversal or
+        bytecode-rewriting sandbox-escape primitives.
+    - field: content
+      operator: regex
+      value: '(?i)CVE-2026-40217[\s\S]{0,200}(?:litellm|guardrail|sandbox|custom[_\s-]?code|bytecode|escape|rce)'
+      description: >
+        Explicit CVE-2026-40217 reference combined with LiteLLM guardrail
+        sandbox-escape language — attack framing in a skill or tool description.
+
+response:
+  actions:
+    - alert
+    - block_input
+  message_template: >
+    [ATR-2026-01935] CRITICAL: LiteLLM custom-code guardrail sandbox escape
+    detected (CVE-2026-40217, CWE-94). Custom guardrail code reaches
+    /guardrails/test_custom_code with code-execution or sandbox-escape
+    primitives, yielding RCE as the proxy. Block the request and upgrade
+    LiteLLM to a build after 2026-04-08.
+
+confidence: 80
+
+wild_fp_rate: 0
+
+test_cases:
+  true_positives:
+    - input: "POST /guardrails/test_custom_code {\"code\": \"import os; os.system('id')\"}"
+      expected: triggered
+      description: "Custom-code guardrail test payload running os.system"
+    - input: "POST /guardrails/test_custom_code {\"code\": \"().__class__.__bases__[0].__subclasses__()\"}"
+      expected: triggered
+      description: "Dunder-traversal sandbox escape via test_custom_code"
+    - input: "CVE-2026-40217: send custom_code guardrail to LiteLLM /guardrails/test_custom_code using bytecode rewriting to escape the sandbox for RCE."
+      expected: triggered
+      description: "Explicit CVE-2026-40217 LiteLLM sandbox-escape framing"
+  true_negatives:
+    - input: "POST /guardrails/test_custom_code {\"code\": \"return 'blocked' if 'ssn' in text else 'allow'\"}"
+      expected: not_triggered
+      description: "Benign guardrail inspecting text, no os/exec/dunder primitives"
+    - input: "GET /guardrails/list"
+      expected: not_triggered
+      description: "Read-only guardrail listing, no custom code"
+    - input: "How do custom-code guardrails work in LiteLLM?"
+      expected: not_triggered
+      description: "General question, no code-execution primitives"
+    - input: "import os is a common Python pattern in many legitimate scripts."
+      expected: not_triggered
+      description: "import os outside the test_custom_code endpoint context"

package/rules/tool-poisoning/ATR-2026-01980-agentic-flow-mcp-tool-param-command-injection.yaml

@@ -0,0 +1,153 @@
+title: "Agentic-Flow MCP Tool-Parameter OS Command Injection (GHSA-vcv2-r9jh-99m5)"
+id: ATR-2026-01980
+rule_version: 1
+status: draft
+description: >
+  Detects GHSA-vcv2-r9jh-99m5 (CWE-78): agentic-flow <= 2.0.13 MCP server
+  tools (agentic_flow_agent, agentic_flow_create_agent, agent_booster_edit_file,
+  agentdb_pattern_store, and related tools under standalone-stdio.ts / fastmcp
+  servers) interpolate attacker-influenceable parameters (agent, task, name,
+  language, agentdb arguments) directly into shell strings passed to
+  execSync(). A parameter value that breaks out of the surrounding
+  double-quoted argument executes arbitrary OS commands with the privileges
+  of the MCP server process.
+
+  Detection covers:
+  (a) an agentic-flow tool-call argument value containing a double-quote
+      breakout followed by shell command chaining (`; `, backticks, or
+      `$(...)`) — the exact PoC shape from the advisory
+      (task = `x"; touch /tmp/INJECTED; id > /tmp/rce.txt; echo "`);
+  (b) the resulting interpolated `npx ... agentic-flow --agent/--task`
+      command string itself carrying injected shell metacharacters;
+  (c) explicit GHSA-vcv2-r9jh-99m5 exploitation framing.
+
+  The detection target is the request shape — a quote-breakout + command
+  chain inside an agentic-flow tool argument — which is the exact
+  command-injection primitive, caught before execSync() hands it to
+  /bin/sh -c. Bound to the agentic-flow tool surface so a benign
+  shell-looking string elsewhere does not fire.
+author: "ATR Community"
+date: "2026/07/04"
+schema_version: "0.1"
+detection_tier: pattern
+maturity: test
+severity: critical
+
+references:
+  owasp_llm:
+    - "LLM06:2025 - Excessive Agency"
+  owasp_agentic:
+    - "ASI06:2026 - Tool Misuse"
+  mitre_atlas:
+    - "AML.T0053 - LLM Plugin Compromise"
+  mitre_attack:
+    - "T1059 - Command and Scripting Interpreter"
+  ghsa:
+    - "GHSA-vcv2-r9jh-99m5"
+
+metadata_provenance:
+  mitre_atlas: human-reviewed
+  owasp_llm: human-reviewed
+  owasp_agentic: human-reviewed
+
+compliance:
+  eu_ai_act:
+    - { article: "15", context: "Article 15 cybersecurity — runtime detection of agentic-flow MCP tool-parameter OS command injection.", strength: primary }
+    - { article: "9",  context: "Article 9 risk management — runtime risk-treatment control for agentic-flow MCP tool-parameter OS command injection.", strength: secondary }
+  nist_ai_rmf:
+    - { subcategory: "MP.5.1", context: "MAP 5.1 — adversarial input characterised/detected for agentic-flow MCP tool-parameter OS command injection.", strength: primary }
+    - { subcategory: "MG.3.2", context: "MANAGE 3.2 — runtime monitoring control for agentic-flow MCP tool-parameter OS command injection.", strength: secondary }
+  iso_42001:
+    - { clause: "8.1", context: "Clause 8.1 operational control — detection of agentic-flow MCP tool-parameter OS command injection payload.", strength: primary }
+    - { clause: "8.3", context: "Clause 8.3 AI risk treatment — runtime detection as treatment control.", strength: secondary }
+
+tags:
+  category: tool-poisoning
+  subcategory: command-injection
+  scan_target: mcp
+  confidence: high
+  source: ghsa-disclosure
+  vendor_sources: agentic-flow-ghsa-vcv2-r9jh-99m5
+
+agent_source:
+  type: tool_call
+  framework:
+    - agentic-flow
+    - any
+  provider:
+    - any
+
+detection:
+  condition: any
+  false_positives:
+    - "A tool_args value that legitimately contains a semicolon or backtick as literal text data (e.g. discussing shell syntax) without an adjacent quote-breakout shape."
+    - "Security advisory or PR text quoting the GHSA-vcv2-r9jh-99m5 payload for review purposes."
+    - "A benign agentic-flow agent/task value with no quote character or shell metacharacter."
+  conditions:
+    - field: tool_args
+      operator: regex
+      value: '(?i)\b(?:agent|task|name|language|agentdb)["\x27]?\s*[:=]\s*["\x27][^"\x27\n]{0,200}["\x27]\s*;\s*[\s\S]{0,120}\b(?:touch|rm|curl|wget|cat|id|nc|bash|sh|chmod|echo)\b'
+      description: >
+        agentic-flow tool argument (agent/task/name/language/agentdb) whose
+        value breaks out of its quote with `";` followed by a chained shell
+        command — the canonical GHSA-vcv2-r9jh-99m5 PoC shape.
+    - field: tool_args
+      operator: regex
+      value: '(?i)npx\s+--yes\s+agentic-flow[\s\S]{0,80}--(?:agent|task)\s+"[^"\n]{0,200}"\s*;\s*[\s\S]{0,120}\b(?:touch|rm|curl|wget|cat|id|nc|bash|sh|chmod)\b'
+      description: >
+        The interpolated `npx --yes agentic-flow --agent/--task` command
+        string itself carrying an injected chained shell command after the
+        closing quote — the post-interpolation execSync() payload.
+    - field: tool_args
+      operator: regex
+      value: '(?i)["\x27][^"\x27\n]{0,200}["\x27]\s*;[\s\S]{0,80}`[^`\n]{0,120}`|["\x27][^"\x27\n]{0,200}["\x27]\s*;[\s\S]{0,80}\$\([^)\n]{0,120}\)'
+      description: >
+        Quote-breakout followed by a backtick or $() command substitution
+        inside an agentic-flow tool argument.
+    - field: content
+      operator: regex
+      value: '(?i)GHSA-vcv2-r9jh-99m5[\s\S]{0,200}(?:agentic-flow|execSync|command\s*inject|RCE)'
+      description: >
+        Explicit GHSA-vcv2-r9jh-99m5 reference combined with agentic-flow
+        command-injection language — attack framing in a skill or tool
+        description.
+
+response:
+  actions:
+    - alert
+    - block_tool
+  message_template: >
+    [ATR-2026-01980] CRITICAL: Agentic-Flow MCP tool-parameter OS command
+    injection detected (GHSA-vcv2-r9jh-99m5, CWE-78). A tool argument breaks
+    out of its shell quote and chains a command; agentic-flow <= 2.0.13
+    interpolates this into execSync() unsanitized. Block the tool call and
+    upgrade agentic-flow to >= 2.0.14.
+
+confidence: 85
+
+wild_fp_rate: 0
+
+test_cases:
+  true_positives:
+    - tool_args: 'task="x\"; touch /tmp/INJECTED; id > /tmp/rce.txt; echo \""'
+      expected: triggered
+      description: "Canonical PoC: task param quote-breakout chaining touch+id+echo"
+    - tool_args: 'npx --yes agentic-flow --agent "coder" --task "x"; touch /tmp/INJECTED; id > /tmp/rce.txt; echo ""'
+      expected: triggered
+      description: "Post-interpolation execSync() command string with injected chain"
+    - tool_description: "Exploiting GHSA-vcv2-r9jh-99m5 in agentic-flow: craft a task value that breaks the execSync quoting for command injection RCE."
+      expected: triggered
+      description: "Explicit GHSA-vcv2-r9jh-99m5 agentic-flow exploitation framing"
+  true_negatives:
+    - tool_args: 'agent="coder", task="Refactor the login handler to use async/await"'
+      expected: not_triggered
+      description: "Legitimate agentic-flow task description, no quote-breakout"
+    - content: "How do I configure the agentic-flow MCP server's agent and task parameters?"
+      expected: not_triggered
+      description: "Benign question about agentic-flow configuration"
+    - tool_args: 'name="release-notes", language="typescript"'
+      expected: not_triggered
+      description: "Normal name/language params with no shell metacharacters"
+    - tool_args: 'agentdb_pattern_search query="find all functions that touch the user; audit them"'
+      expected: not_triggered
+      description: "Benign text containing the word touch and a semicolon in natural language, no quote-breakout or dangerous command token"

package/rules/tool-poisoning/ATR-2026-01982-dbt-mcp-node-selection-argument-injection.yaml

@@ -0,0 +1,160 @@
+title: "dbt-mcp node_selection/resource_type Argument Injection (CVE-2026-44968)"
+id: ATR-2026-01982
+rule_version: 1
+status: draft
+description: >
+  Detects CVE-2026-44968 / GHSA-xpww-f6pm-cfhq (CWE-88): dbt-mcp's
+  _run_dbt_command() in src/dbt_mcp/dbt_cli/tools.py appends the
+  MCP-client-supplied node_selection string (split on space) and
+  resource_type JSON array verbatim to the dbt subprocess argv without
+  validating that tokens don't begin with a dash. Because Popen is called
+  with shell=False, shell metacharacters are inert, but an attacker can still
+  inject dbt global flags — --profiles-dir, --project-dir, --target,
+  --profile — as independent argv elements on the build/compile/run/test/
+  clone/list/get_node_details_dev tools, redirecting dbt's configuration,
+  project root, or execution target (demonstrated PoC: node_selection =
+  "my_model --profiles-dir /tmp/evil" loads an attacker-controlled
+  profiles.yml and writes to an attacker-chosen database path).
+
+  Detection covers:
+  (a) a node_selection value containing an injected dbt global flag token
+      after the legitimate selector (space-separated, so the flag rides
+      along as an extra "word");
+  (b) a resource_type JSON array containing an injected flag token as an
+      array element instead of a valid resource type;
+  (c) explicit CVE-2026-44968 / GHSA-xpww-f6pm-cfhq exploitation framing.
+
+  The detection target is the request shape — a dash-prefixed dbt global
+  flag riding inside a selector/resource-type value that should only ever
+  contain a selector or resource-type token — which is the exact argument-
+  injection primitive, caught before _run_dbt_command() extends argv. Bound
+  to the dbt-mcp tool surface (node_selection/resource_type params) so a
+  benign selector string elsewhere does not fire.
+author: "ATR Community"
+date: "2026/07/04"
+schema_version: "0.1"
+detection_tier: pattern
+maturity: test
+severity: medium
+
+references:
+  owasp_llm:
+    - "LLM06:2025 - Excessive Agency"
+  owasp_agentic:
+    - "ASI06:2026 - Tool Misuse"
+  mitre_atlas:
+    - "AML.T0053 - LLM Plugin Compromise"
+  mitre_attack:
+    - "T1059 - Command and Scripting Interpreter"
+  cve:
+    - "CVE-2026-44968"
+  ghsa:
+    - "GHSA-xpww-f6pm-cfhq"
+
+metadata_provenance:
+  mitre_atlas: human-reviewed
+  owasp_llm: human-reviewed
+  owasp_agentic: human-reviewed
+
+compliance:
+  eu_ai_act:
+    - { article: "15", context: "Article 15 cybersecurity — runtime detection of dbt-mcp node_selection/resource_type argument injection.", strength: primary }
+    - { article: "9",  context: "Article 9 risk management — runtime risk-treatment control for dbt-mcp node_selection/resource_type argument injection.", strength: secondary }
+  nist_ai_rmf:
+    - { subcategory: "MP.5.1", context: "MAP 5.1 — adversarial input characterised/detected for dbt-mcp node_selection/resource_type argument injection.", strength: primary }
+    - { subcategory: "MG.3.2", context: "MANAGE 3.2 — runtime monitoring control for dbt-mcp node_selection/resource_type argument injection.", strength: secondary }
+  iso_42001:
+    - { clause: "8.1", context: "Clause 8.1 operational control — detection of dbt-mcp node_selection/resource_type argument injection payload.", strength: primary }
+    - { clause: "8.3", context: "Clause 8.3 AI risk treatment — runtime detection as treatment control.", strength: secondary }
+
+tags:
+  category: tool-poisoning
+  subcategory: argument-injection
+  scan_target: mcp
+  confidence: high
+  source: cve-disclosure
+  vendor_sources: dbt-mcp-cve-2026-44968
+
+agent_source:
+  type: tool_call
+  framework:
+    - dbt-mcp
+    - any
+  provider:
+    - any
+
+detection:
+  condition: any
+  false_positives:
+    - "A node_selection value that legitimately contains a dbt selector method using a colon (tag:, config:) rather than a dash-prefixed flag."
+    - "Security advisory or PR text quoting the CVE-2026-44968 / GHSA-xpww-f6pm-cfhq payload for review purposes."
+    - "A resource_type array containing only valid dbt resource type names (model, test, snapshot, source, etc.) with no injected flag."
+  conditions:
+    - field: tool_args
+      operator: regex
+      value: '(?i)\bnode_selection["\x27]?\s*[:=]\s*["\x27][^"\x27\n]{0,200}\s(?:--profiles-dir|--project-dir|--target|--profile)\b'
+      description: >
+        node_selection value whose space-separated tokens include an
+        injected dbt global flag (--profiles-dir/--project-dir/--target/
+        --profile) after the legitimate selector — CVE-2026-44968 Vector 1.
+    - field: tool_args
+      operator: regex
+      value: '(?i)\bresource_type["\x27]?\s*[:=]\s*\[[^\]\n]{0,200}["\x27](?:--profiles-dir|--project-dir|--target|--profile)["\x27][^\]\n]{0,200}\]'
+      description: >
+        resource_type JSON array containing an injected dbt global flag
+        token as an array element instead of a valid resource type —
+        CVE-2026-44968 Vector 2.
+    - field: tool_args
+      operator: regex
+      value: '(?i)--(?:profiles-dir|project-dir|target|profile)\s+["\x27]?/tmp/[^"\x27\s\n]{0,80}["\x27]?[\s\S]{0,120}\b(?:node_selection|resource_type|--select|--resource-type)\b'
+      description: >
+        An injected dbt global flag pointing at an attacker-controlled path
+        appearing alongside the vulnerable --select/--resource-type argv
+        construction.
+    - field: content
+      operator: regex
+      value: '(?i)(?:CVE-2026-44968|GHSA-xpww-f6pm-cfhq)[\s\S]{0,200}(?:dbt-mcp|node_selection|resource_type|profiles-dir|argument\s*inject)'
+      description: >
+        Explicit CVE-2026-44968 / GHSA-xpww-f6pm-cfhq reference combined with
+        dbt-mcp argument-injection language — attack framing in a skill or
+        tool description.
+
+response:
+  actions:
+    - alert
+    - block_tool
+  message_template: >
+    [ATR-2026-01982] MEDIUM: dbt-mcp node_selection/resource_type argument
+    injection detected (CVE-2026-44968, CWE-88). A tool argument injects a
+    dbt global flag (--profiles-dir/--project-dir/--target) as an argv
+    element; dbt-mcp before 1.17.1 appends this to the dbt subprocess
+    unsanitized. Block the tool call and upgrade dbt-mcp to >= 1.17.1.
+
+confidence: 80
+
+wild_fp_rate: 0
+
+test_cases:
+  true_positives:
+    - tool_args: 'node_selection="my_first_model --profiles-dir /tmp/evil-profiles"'
+      expected: triggered
+      description: "PoC Vector 1: node_selection injects --profiles-dir pointing at attacker path"
+    - tool_args: 'resource_type=["model", "--profiles-dir", "/tmp/evil-profiles"]'
+      expected: triggered
+      description: "PoC Vector 2: resource_type array smuggles --profiles-dir as an element"
+    - tool_description: "Exploiting CVE-2026-44968 (GHSA-xpww-f6pm-cfhq) in dbt-mcp: craft node_selection to inject --profiles-dir for argument injection against the dbt CLI."
+      expected: triggered
+      description: "Explicit CVE-2026-44968 dbt-mcp exploitation framing"
+  true_negatives:
+    - tool_args: 'node_selection="my_first_model+ tag:nightly"'
+      expected: not_triggered
+      description: "Legitimate dbt selector syntax with plus-operator and tag: method, no injected flag"
+    - tool_args: 'resource_type=["model", "test", "source"]'
+      expected: not_triggered
+      description: "Valid resource_type array with only real dbt resource type names"
+    - content: "How do I filter dbt models using node_selection in the dbt-mcp run tool?"
+      expected: not_triggered
+      description: "Benign question about dbt-mcp node_selection usage"
+    - tool_args: 'node_selection="staging.orders"'
+      expected: not_triggered
+      description: "Simple model selector, no flags, read-only compile intent"

package/rules/tool-poisoning/ATR-2026-01983-mcp-for-stata-log-file-name-command-injection.yaml

@@ -0,0 +1,170 @@
+title: "MCP-for-Stata: Command Injection via log_file_name Parameter (CVE-2026-47708)"
+id: ATR-2026-01983
+rule_version: 1
+status: draft
+description: >
+  Detects CVE-2026-47708 (CVSS critical, CWE-77): the `log_file_name` (aka
+  `log_name`) parameter accepted by the `stata_do` MCP tool / CLI is
+  interpolated directly into a Stata `log using "<log_file>", ...` command
+  string without sanitization. GuardValidator only scans do-file content, not
+  this wrapper parameter. An attacker breaks out of the quoted log-path
+  string with a quote or newline and appends a dangerous Stata command
+  (`shell`, `python`, `erase`, `winexec`, or `!` shell-escape), achieving
+  remote code execution, or supplies `../` path-traversal segments in
+  log_name to write arbitrary files outside the log directory.
+
+  Detection covers:
+  (a) a stata_do tool call whose log_file_name/log_name value contains a
+      quote/newline breakout followed by a dangerous Stata command token;
+  (b) a log_file_name/log_name value containing path-traversal segments;
+  (c) explicit CVE-2026-47708 exploitation framing referencing the
+      log_file_name / stata_do injection.
+
+  The detection target is the request shape — a stata_do log_file_name
+  parameter carrying a string-breakout plus command injection or path
+  traversal — the exact CVE-2026-47708 primitive, caught before the
+  wrapper builds the Stata command string.
+author: "ATR Community"
+date: "2026/07/04"
+schema_version: "0.1"
+detection_tier: pattern
+maturity: test
+severity: critical
+
+references:
+  owasp_llm:
+    - "LLM06:2025 - Excessive Agency"
+  owasp_agentic:
+    - "ASI06:2026 - Tool Misuse"
+  mitre_atlas:
+    - "AML.T0053 - LLM Plugin Compromise"
+  mitre_attack:
+    - "T1059 - Command and Scripting Interpreter"
+  cve:
+    - "CVE-2026-47708"
+
+metadata_provenance:
+  mitre_atlas: human-reviewed
+  owasp_llm: human-reviewed
+  owasp_agentic: human-reviewed
+
+compliance:
+  eu_ai_act:
+    - article: "15"
+      context: "Article 15 cybersecurity — runtime detection of MCP-for-Stata log_file_name command injection (CVE-2026-47708)."
+      strength: primary
+    - article: "9"
+      context: "Article 9 risk management — runtime risk-treatment control for MCP-for-Stata log_file_name command injection (CVE-2026-47708)."
+      strength: secondary
+  nist_ai_rmf:
+    - subcategory: "MP.5.1"
+      context: "MAP 5.1 — adversarial input characterised/detected for MCP-for-Stata log_file_name command injection (CVE-2026-47708)."
+      strength: primary
+    - subcategory: "MG.3.2"
+      context: "MANAGE 3.2 — runtime monitoring control for MCP-for-Stata log_file_name command injection (CVE-2026-47708)."
+      strength: secondary
+  iso_42001:
+    - clause: "8.1"
+      context: "Clause 8.1 operational control — detection of MCP-for-Stata log_file_name command injection payload."
+      strength: primary
+    - clause: "8.3"
+      context: "Clause 8.3 AI risk treatment — runtime detection as treatment control."
+      strength: secondary
+
+tags:
+  category: tool-poisoning
+  subcategory: mcp-command-injection
+  scan_target: runtime
+  confidence: high
+  source: cve-disclosure
+  vendor_sources: mcp-for-stata-cve-2026-47708
+
+agent_source:
+  type: tool_call
+  framework:
+    - any
+  provider:
+    - any
+
+detection:
+  condition: any
+  false_positives:
+    - "A log_file_name value that legitimately contains a single quote as part of a project naming convention but no command token follows."
+    - "Security advisory or PR text quoting the CVE-2026-47708 payload."
+    - "A stata_do call with a plain alphanumeric log_file_name and no quote/newline breakout."
+  conditions:
+    - field: tool_args
+      operator: regex
+      value: '(?i)\b(?:log_file_name|log_name)\b["\x27]?\s*[:=]\s*["\x27][\s\S]{0,80}["\x27][\s\S]{0,40}(?:\bshell\b|\bpython\b|\berase\b|\bwinexec\b|(?<![\w=])!(?!=))'
+      description: >
+        A stata_do log_file_name/log_name parameter value that breaks out of
+        its quoted string and is followed by a dangerous Stata command token
+        (shell / python / erase / winexec / ! shell-escape) — the
+        CVE-2026-47708 command-injection breakout shape.
+    - field: tool_args
+      operator: regex
+      value: '(?i)\blog\s+using\s+["\x27][\s\S]{0,120}["\x27][\s\S]{0,40}(?:\bshell\b|\bpython\b|\berase\b|\bwinexec\b)\b'
+      description: >
+        A generated Stata `log using "<path>", ...` command string where the
+        interpolated log path breaks out and chains into a dangerous Stata
+        command — the exact string the vulnerable generate_log_command
+        produces once log_name is attacker-controlled.
+    - field: tool_args
+      operator: regex
+      value: '(?i)\b(?:log_file_name|log_name)\b["\x27]?\s*[:=]\s*["\x27][\s\S]{0,20}(?:\.\.[\\/]){2,}'
+      description: >
+        A stata_do log_file_name/log_name parameter containing repeated
+        path-traversal segments (../) — arbitrary file write outside the
+        intended log directory.
+    - field: tool_args
+      operator: regex
+      value: '(?i)CVE-2026-47708[\s\S]{0,200}(?:log_file_name|log_name|stata_do|GuardValidator)'
+      description: >
+        Explicit CVE-2026-47708 reference combined with the vulnerable
+        stata_do log_file_name/log_name parameter or GuardValidator guard —
+        attack framing in a skill or tool description.
+
+response:
+  actions:
+    - alert
+    - block_tool
+  message_template: >
+    [ATR-2026-01983] CRITICAL: MCP-for-Stata log_file_name command injection
+    detected (CVE-2026-47708, CWE-77). The stata_do log_file_name/log_name
+    parameter breaks out of its quoted Stata command string and injects a
+    dangerous command (shell/python/erase/winexec) or performs path
+    traversal. GuardValidator only scans do-file content, not this
+    parameter. Block the tool call and apply allowlist validation
+    (alphanumeric/underscore/dot/hyphen only) to log_name.
+
+confidence: 82
+
+wild_fp_rate: 0
+
+test_cases:
+  true_positives:
+    - tool_args: "stata_do {\"dofile_path\": \"test.do\", \"log_file_name\": \"'; shell echo pwned > /tmp/pwned.txt; '\"}"
+      expected: triggered
+      description: "CVE-2026-47708 PoC — log_file_name breaks out of quoted string and injects a shell command"
+    - tool_args: 'log using "/var/log/stata/x", replace text name(text_log) shell rm -rf ~'
+      expected: triggered
+      description: "Generated Stata log command with breakout chaining into a shell command"
+    - tool_args: 'log_file_name: "../../../../etc/cron.d/pwned"'
+      expected: triggered
+      description: "log_file_name path traversal to write outside the intended log directory"
+    - tool_args: "Exploiting CVE-2026-47708 in mcp-for-stata: craft a malicious log_file_name to break out of the quoted log using command and reach GuardValidator-unchecked stata_do."
+      expected: triggered
+      description: "Explicit CVE-2026-47708 mcp-for-stata exploitation framing"
+  true_negatives:
+    - tool_args: 'stata_do {"dofile_path": "analysis.do", "log_file_name": "session_2026_07_04"}'
+      expected: not_triggered
+      description: "Legitimate stata_do call with a plain alphanumeric log_file_name"
+    - tool_args: 'log using "/var/log/stata/session_2026_07_04.log", replace text name(text_log)'
+      expected: not_triggered
+      description: "Legitimate generated log using command with no breakout or injected command"
+    - tool_args: "How do I set a custom log_file_name when calling stata_do?"
+      expected: not_triggered
+      description: "General question about the log_file_name parameter, not an attack payload"
+    - tool_args: 'stata_do {"dofile_path": "regress.do", "log_name": "project-alpha_v2"}'
+      expected: not_triggered
+      description: "Read-only-style legitimate call with hyphenated project log name, no quote breakout"