{
  "schema_version": "1.6.0",
  "id": "ACVE-2026-0203",
  "aliases": [],
  "published": "2025-09-26T00:00:00Z",
  "firstReported": {
    "date": "2025-09-26",
    "url": "https://arxiv.org/abs/2509.22830",
    "publisher": "ChatInject (arXiv)"
  },
  "modified": "2026-09-25T00:00:00Z",
  "summary": "A tool-result injection in the model's own chat-template role tokens is obeyed as a system or user turn by auto-approved agents on Qwen3-235B-A22B, GPT-oss-120b, GLM-4.5, Llama-4-Maverick or Kimi-K2.",
  "details": "## What\n\nChatInject (arXiv:2509.22830) wraps an injection in the target model's chat-template role tokens, so text inside a tool result reads as a higher-priority turn. Table 1: AgentDojo ASR rose from 17.5% to 54.8% on Qwen3-235B-A22B, 0.3% to 51.4% on GPT-oss-120b, 0.3% to 20.3% on GLM-4.5 and 1.0% to 17.2% on Llama-4-Maverick; the multi-turn variant reached 80.5% on Qwen3. Tools are simulated. It amplifies injection but needs the ACVE-2026-0201 exposure to cause harm.\n\n## Detection\n\nThe lockfile sees the model family, the tool attributes and the approval mode; it cannot see whether tool output is escaped before it enters the template.\n\n## Fix\n\nRequire approval for side effects. Escaping role tokens in tool output helps, but the paper warns \"deterministic format filters alone are insufficient\": with 10% of template characters perturbed the attack still beat the plain-text baseline.\n",
  "affected": [
    {
      "package": {
        "ecosystem": "Model",
        "name": "llama-4"
      },
      "ranges": [
        {
          "type": "ECOSYSTEM",
          "events": [
            {
              "introduced": "0"
            }
          ]
        }
      ]
    },
    {
      "package": {
        "ecosystem": "Model",
        "name": "qwen3"
      },
      "ranges": [
        {
          "type": "ECOSYSTEM",
          "events": [
            {
              "introduced": "0"
            }
          ]
        }
      ]
    },
    {
      "package": {
        "ecosystem": "Model",
        "name": "glm-4.5"
      },
      "ranges": [
        {
          "type": "ECOSYSTEM",
          "events": [
            {
              "introduced": "0"
            }
          ]
        }
      ]
    },
    {
      "package": {
        "ecosystem": "Model",
        "name": "gpt-oss-120b"
      },
      "ranges": [
        {
          "type": "ECOSYSTEM",
          "events": [
            {
              "introduced": "0"
            }
          ]
        }
      ]
    },
    {
      "package": {
        "ecosystem": "Model",
        "name": "kimi-k2"
      },
      "ranges": [
        {
          "type": "ECOSYSTEM",
          "events": [
            {
              "introduced": "0"
            }
          ]
        }
      ]
    }
  ],
  "references": [
    {
      "type": "ARTICLE",
      "url": "https://arxiv.org/abs/2509.22830"
    },
    {
      "type": "EVIDENCE",
      "url": "https://arxiv.org/html/2509.22830"
    },
    {
      "type": "EXPLOIT",
      "url": "https://hwanchang00.github.io/chatinject_project_page/"
    }
  ],
  "database_specific": {
    "severity": "MODERATE",
    "acve": {
      "specVersion": 1,
      "status": "candidate",
      "kind": "behavioural",
      "vulnClasses": [
        "prompt-injection-to-tool"
      ],
      "cwe": [
        "CWE-1427"
      ],
      "taxonomy": {
        "owasp_llm": [
          "LLM01"
        ],
        "owasp_asi": [
          "ASI01",
          "ASI02"
        ],
        "atlas": [
          "AML.T0051.001",
          "AML.T0065"
        ]
      },
      "noCveReason": "attack technique against model chat templates; no code defect in one product",
      "threat": {
        "attacker": "content-author",
        "vector": "prompt-injection",
        "outcome": "harmful-action"
      },
      "evidence": [
        {
          "kind": "eval",
          "benchmark": "AgentDojo",
          "attack": "InjecPrompt + ChatInject",
          "defense": "none",
          "metric": "asr",
          "value": 0.548,
          "k": 1,
          "ci95": [
            0.498,
            0.596
          ],
          "model": {
            "id": "Qwen3-235B-A22B"
          },
          "url": "https://arxiv.org/html/2509.22830",
          "note": "Table 1; Wilson 95% CI from Table 10; simulated tools; temperature 0 via OpenRouter"
        },
        {
          "kind": "eval",
          "benchmark": "AgentDojo",
          "attack": "InjecPrompt + ChatInject",
          "defense": "none",
          "metric": "asr",
          "value": 0.514,
          "k": 1,
          "ci95": [
            0.465,
            0.563
          ],
          "model": {
            "id": "GPT-oss-120b"
          },
          "url": "https://arxiv.org/html/2509.22830",
          "note": "Table 1; Wilson 95% CI from Table 10; simulated tools; temperature 0 via OpenRouter"
        },
        {
          "kind": "eval",
          "benchmark": "AgentDojo",
          "attack": "InjecPrompt + ChatInject",
          "defense": "none",
          "metric": "asr",
          "value": 0.203,
          "k": 1,
          "ci95": [
            0.166,
            0.246
          ],
          "model": {
            "id": "GLM-4.5"
          },
          "url": "https://arxiv.org/html/2509.22830",
          "note": "Table 1; Wilson 95% CI from Table 10; simulated tools; temperature 0 via OpenRouter"
        },
        {
          "kind": "eval",
          "benchmark": "AgentDojo",
          "attack": "InjecPrompt + ChatInject",
          "defense": "none",
          "metric": "asr",
          "value": 0.172,
          "k": 1,
          "ci95": [
            0.138,
            0.213
          ],
          "model": {
            "id": "Llama-4-Maverick"
          },
          "url": "https://arxiv.org/html/2509.22830",
          "note": "Table 1; Wilson 95% CI from Table 10; simulated tools; temperature 0 via OpenRouter"
        },
        {
          "kind": "eval",
          "benchmark": "InjecAgent",
          "attack": "InjecPrompt + ChatInject",
          "defense": "none",
          "metric": "asr",
          "value": 0.794,
          "k": 1,
          "ci95": [
            0.769,
            0.817
          ],
          "model": {
            "id": "Llama-4-Maverick"
          },
          "url": "https://arxiv.org/html/2509.22830",
          "note": "Table 1; Wilson 95% CI from Table 10; simulated tools; temperature 0 via OpenRouter"
        },
        {
          "kind": "eval",
          "benchmark": "InjecAgent",
          "attack": "InjecPrompt + ChatInject",
          "defense": "none",
          "metric": "asr",
          "value": 0.573,
          "k": 1,
          "ci95": [
            0.543,
            0.603
          ],
          "model": {
            "id": "GLM-4.5"
          },
          "url": "https://arxiv.org/html/2509.22830",
          "note": "Table 1; Wilson 95% CI from Table 10; simulated tools; temperature 0 via OpenRouter"
        },
        {
          "kind": "eval",
          "benchmark": "InjecAgent",
          "attack": "InjecPrompt + ChatInject",
          "defense": "none",
          "metric": "asr",
          "value": 0.394,
          "k": 1,
          "ci95": [
            0.365,
            0.424
          ],
          "model": {
            "id": "Qwen3-235B-A22B"
          },
          "url": "https://arxiv.org/html/2509.22830",
          "note": "Table 1; Wilson 95% CI from Table 10; simulated tools; temperature 0 via OpenRouter"
        },
        {
          "kind": "eval",
          "benchmark": "InjecAgent",
          "attack": "InjecPrompt + ChatInject",
          "defense": "none",
          "metric": "asr",
          "value": 0.142,
          "k": 1,
          "ci95": [
            0.123,
            0.165
          ],
          "model": {
            "id": "GPT-oss-120b"
          },
          "url": "https://arxiv.org/html/2509.22830",
          "note": "Table 1; Wilson 95% CI from Table 10; simulated tools; temperature 0 via OpenRouter"
        }
      ],
      "matcher": {
        "all": [
          {
            "any": [
              {
                "all": [
                  {
                    "field": "model.id",
                    "op": "exists"
                  },
                  {
                    "field": "model.family",
                    "op": "in",
                    "value": [
                      "llama-4",
                      "qwen3",
                      "glm-4.5",
                      "gpt-oss",
                      "gpt-oss-120b",
                      "kimi-k2"
                    ]
                  }
                ]
              },
              {
                "some": "models",
                "where": {
                  "field": "family",
                  "op": "in",
                  "value": [
                    "llama-4",
                    "qwen3",
                    "glm-4.5",
                    "gpt-oss",
                    "gpt-oss-120b",
                    "kimi-k2"
                  ]
                }
              }
            ]
          },
          {
            "some": "tools",
            "where": {
              "field": "untrustedInput",
              "op": "eq",
              "value": true
            }
          },
          {
            "some": "tools",
            "where": {
              "field": "sideEffect",
              "op": "eq",
              "value": true
            }
          },
          {
            "field": "approval",
            "op": "in",
            "value": [
              "never",
              "allowlist"
            ]
          }
        ]
      },
      "fix": {
        "summary": "Require approval for side effects; escape role tokens in tool output, which the paper shows is not sufficient by itself.",
        "actions": [
          {
            "type": "reconfigure",
            "target": "agent.approval",
            "to": "ask",
            "why": "Format filtering is bypassable, so the approval boundary is the control that holds.",
            "owner": "operator"
          },
          {
            "type": "reconfigure",
            "target": "harness.toolOutput",
            "to": "chat-template role tokens escaped before tool output enters the template",
            "why": "Recommended: the attack forges system and user turns with role tokens; the paper finds filtering alone insufficient.",
            "owner": "harness-vendor"
          }
        ]
      },
      "exposure": {
        "harness": {
          "value": "Any harness that inserts tool output into the model's chat template unescaped; the paper ran the AgentDojo and InjecAgent harnesses against the models through OpenRouter at temperature 0",
          "any": true,
          "status": "confirmed",
          "source": "https://arxiv.org/html/2509.22830"
        },
        "model": {
          "value": "Qwen3-235B-A22B, GPT-oss-120b, GLM-4.5, Llama-4-Maverick and Kimi-K2 (the paper also tests GPT-4o, Grok-2, Grok-3, Gemini-2.5-Pro and Gemma-3)",
          "any": false,
          "status": "confirmed",
          "source": "https://arxiv.org/html/2509.22830"
        },
        "goal": {
          "value": "An AgentDojo user task in the banking, Slack or travel-booking suite, or an InjecAgent user task, whose tool results the attacker can write into",
          "stated": true,
          "status": "confirmed",
          "source": "https://arxiv.org/html/2509.22830"
        },
        "tools": {
          "value": "Simulated AgentDojo and InjecAgent tools whose results are inserted into the model's chat template unescaped; the model's public chat-template role tokens",
          "status": "confirmed",
          "source": "https://arxiv.org/html/2509.22830"
        },
        "approval": {
          "value": "The benchmark harnesses execute every tool call the model emits; the paper describes no confirmation step",
          "mode": "auto-approve",
          "status": "confirmed",
          "source": "https://arxiv.org/html/2509.22830"
        },
        "inputControl": "content-author",
        "agentAction": "The agent treats the forged system or user turn inside the tool result as an instruction and calls the tool the injection asked for.",
        "harm": "harmful-action",
        "divergence": "instruction-followed",
        "reach": {
          "value": "The simulated tools of the benchmark environments",
          "kinds": [
            "network"
          ],
          "status": "confirmed",
          "source": "https://arxiv.org/abs/2509.22830"
        },
        "condition": {
          "value": "When tool output is inserted into the chat template without escaping role tokens",
          "status": "confirmed",
          "source": "https://arxiv.org/abs/2509.22830"
        }
      },
      "occurrence": {
        "basis": "demonstrated",
        "reportedBy": "researcher"
      },
      "exploitation": {
        "status": "demonstrated",
        "checkedAt": "2026-09-21T00:00:00Z",
        "sources": [
          {
            "url": "https://arxiv.org/abs/2509.22830",
            "type": "research",
            "note": "benchmark demonstration on simulated tools"
          }
        ],
        "kev": {
          "listed": false
        }
      },
      "cveBoundary": "behavioural",
      "reproducibility": {
        "status": "partial",
        "axesComplete": true,
        "componentsObtainable": null,
        "triggerPublished": true,
        "observableStated": true,
        "demonstrated": "researcher-demonstrated",
        "missing": [
          "model availability not checked (no registry reference)"
        ]
      },
      "severityBasis": "eval-asr"
    }
  }
}
