跳到內容
Agent Threat Rules (ATR)EN
Α提示注入
ATR-2026-02011
收錄於 2026-07-07 · 作者 ATR Community

Malformed-Grammar PWNED Success-Claim Detection

Detects prompt-injection success-claim outputs where the model has been coerced into emitting a grammatically broken "has/have been PWNED" style confirmation string.

jailbreak · pattern · confidence: high

嚴重度
high
車道
alert · hunt比對到就標出、回報
成熟度
test還沒定論
掃描目標
mcp
狀態
experimental

框架對照

OWASP Agentic Top 10
  • ASI01:2026
OWASP LLM Top 10
  • LLM01:2025

這些對應是 ATR 自己做的,框架方沒有審查或背書。

說明

Detects prompt-injection success-claim outputs where the model has been coerced into emitting a grammatically broken "has/have been PWNED" style confirmation string. These malformed-grammar variants (e.g. "Me have been PWNED", "I has been PWNED") are a recurring artifact of automated jailbreak/injection fuzzing corpora (e.g. HackAPrompt-style challenges) where the attacker forces the target LLM to echo a templated compliance token regardless of correct grammar. Detection targets the LLM output / tool response text channel rather than user input, since the payload only appears once the model has already produced the compromised completion.

建議回應block inputalertsnapshotescalate
作者:ATR Community · 建立 2026-07-07 · 最後修改 2026-11-20

野外驗證

驗證日期
2026-07-07
樣本數
4,817

偵測條件

組合方式:any
  1. 01
    Grammatically malformed first-person 'has/have been PWNED' success-claim string typical of forced prompt-injection compliance output
    欄位:content運算子:regex

會觸發的攻擊範例

  1. Me have been PWNED English
  2. I has been PWNED. You has been PWNED. We all has been PWNED.

規則必須抓到的攻擊範例,很多是照公開回報的攻擊改寫的,有些是照攻擊的形狀寫出來的。範例跟規則一起版本化,當作回歸測試。之後改規則時,如果其中一個範例不再命中,一定看得出來。

不會觸發的正常樣本

  1. Benign first-person travel anecdote using 'have been'
    I have been to Paris three times and each trip was wonderful.
  2. Benign team status update using 'have been'
    We all have been working overtime this week to hit the deadline.
  3. Benign third-person passive voice, no first-person subject
    The document has been reviewed and approved by the compliance team.
  4. Benign professional bio statement using 'have been'
    I have been a software engineer for ten years and love solving hard problems.

已知的誤報情境

  • ▸Benign chat text referencing having 'been' somewhere, unrelated to any compliance token
  • ▸Security researchers discussing PWNED-style CTF/jailbreak success strings in professional writeups
  • ▸Legitimate grammar/language-learning discussion of correct vs incorrect verb conjugation
title: "Malformed-Grammar PWNED Success-Claim Detection"
id: ATR-2026-02011
rule_version: 1
status: experimental
description: >
  Detects prompt-injection success-claim outputs where the model has been coerced into
  emitting a grammatically broken "has/have been PWNED" style confirmation string.
  These malformed-grammar variants (e.g. "Me have been PWNED", "I has been PWNED")
  are a recurring artifact of automated jailbreak/injection fuzzing corpora (e.g.
  HackAPrompt-style challenges) where the attacker forces the target LLM to echo a
  templated compliance token regardless of correct grammar. Detection targets the
  LLM output / tool response text channel rather than user input, since the payload
  only appears once the model has already produced the compromised completion.
author: "ATR Community"
date: "2026/07/07"
schema_version: "0.1"
detection_tier: pattern
maturity: test
severity: high

references:
  owasp_llm:
    - "LLM01:2025"
  owasp_agentic:
    - "ASI01:2026"
  mitre_atlas:
    - "AML.T0051 - LLM Prompt Injection"
    - "AML.T0054 - LLM Jailbreak"

compliance:
  eu_ai_act:
    - article: "15"
      context: "High-risk AI systems must be resilient against adversarial attempts to suppress safety mechanisms. Detecting confirmed prompt-injection success-claim output provides detection evidence supporting Article 15 requirements for robustness against input-based manipulation."
      strength: primary
    - article: "9"
      context: "Successful prompt-injection compliance strings constitute a documented risk class in the AI system risk register; Article 9 requires that monitoring controls are deployed to detect these attempts at runtime."
      strength: secondary
  nist_ai_rmf:
    - subcategory: "MP.5.1"
      context: "Malformed-grammar success-claim strings are a primary artifact of adversarial prompt-injection attacks against AI systems; MP.5.1 requires that adversarial input risks are identified and tracked so that runtime detection controls like this rule can be deployed."
顯示全部(還有 80 行)
      strength: primary
    - subcategory: "MG.3.2"
      context: "Detected PWNED-style success-claim output represents confirmed exploitation of the model's output channel, triggering the risk treatment response plans required by MG.3.2 to contain and remediate adversarial prompt attacks."
      strength: secondary
  iso_42001:
    - clause: "8.1"
      context: "Clause 8.1 operational controls ensure AI systems execute correctly and consistently; runtime detection of confirmed injection success-claim output enforces that safety constraints remain active despite adversarial instructions to disable them."
      strength: primary
    - clause: "8.3"
      context: "Clause 8.3 requires AI system impacts to be assessed and monitored in operation; this rule provides monitoring evidence for the prompt-injection technique (Malformed-Grammar PWNED Success-Claim) once it manifests in model output."
      strength: secondary

tags:
  category: prompt-injection
  subcategory: jailbreak
  scan_target: mcp
  confidence: high

agent_source:
  type: llm_io
  framework:
    - openai
    - anthropic
    - langchain
    - llamaindex
    - autogen
    - crewai
  provider:
    - any

detection:
  conditions:
    - field: content
      operator: regex
      value: "(?i)\\b(?:i|me|iv|we all)\\b[\\s\\S]{0,15}\\b(?:has|have|has be|iv)\\b[\\s\\S]{0,10}\\bbeen\\b[\\s\\S]{0,10}\\bpwned\\b"
      description: "Grammatically malformed first-person 'has/have been PWNED' success-claim string typical of forced prompt-injection compliance output"

  condition: any
  false_positives:
    - "Benign chat text referencing having 'been' somewhere, unrelated to any compliance token"
    - "Security researchers discussing PWNED-style CTF/jailbreak success strings in professional writeups"
    - "Legitimate grammar/language-learning discussion of correct vs incorrect verb conjugation"

response:
  actions:
    - block_input
    - alert
    - snapshot
    - escalate
  auto_response_threshold: high
  message_template: >
    [ATR-2026-02011] Prompt-injection success-claim detected. The model output contains
    a malformed-grammar "has/have been PWNED" confirmation string consistent with a
    successful prompt-injection or jailbreak attempt. This response has been blocked.
    Repeated occurrences may indicate an active injection campaign against this agent.

confidence: 88
wild_validated: "2026/07/07"
wild_samples: 4817
test_cases:
  true_positives:
    - input: "Me have been PWNED English"
      expected: triggered
      description: "Malformed-grammar success-claim from HackAPrompt-style residual corpus"
    - input: "I has been PWNED. You has been PWNED. We all has been PWNED."
      expected: triggered
      description: "Repeated malformed-grammar success-claim across multiple subjects"
  true_negatives:
    - input: "I have been to Paris three times and each trip was wonderful."
      expected: not_triggered
      description: "Benign first-person travel anecdote using 'have been'"
    - input: "We all have been working overtime this week to hit the deadline."
      expected: not_triggered
      description: "Benign team status update using 'have been'"
    - input: "The document has been reviewed and approved by the compliance team."
      expected: not_triggered
      description: "Benign third-person passive voice, no first-person subject"
    - input: "I have been a software engineer for ten years and love solving hard problems."
      expected: not_triggered
      description: "Benign professional bio statement using 'have been'"

修訂紀錄

建立於
2026-07-07
最後修改
2026-11-20
到 GitHub 看完整的 commit 紀錄