{
  "schema_version": "1.6.0",
  "id": "ACVE-2026-0218",
  "aliases": [],
  "published": "2026-07-24T00:00:00Z",
  "firstReported": {
    "date": "2026-07-24",
    "url": "https://www-cdn.anthropic.com/c5fbac3f0b1280a933ebd26d3cb8bb9f5bdeaf48/Claude%20Opus%205%20System%20Card.pdf",
    "publisher": "Anthropic (Claude Opus 5 system card)"
  },
  "modified": "2026-09-25T00:00:00Z",
  "summary": "Indirect injections drive agents on GPT 5.6 Terra or GPT 5.6 Luna, with side effects auto-approved, into irreversible actions at rates the Claude Opus 5 System Card measured well above GPT 5.6 Sol.",
  "details": "## What\n\nThe Claude Opus 5 System Card (Anthropic, 2026-07-24, section 5.2.1, Figure 5.2.1.B) reports the Gray Swan IPI benchmark: the probability of at least one successful indirect injection within 15 attempts was 30.4% for GPT 5.6 Terra and 43.9% for Luna, against 20.0% for Sol (3.1% at one attempt) and 20.8% for GPT 5.5. The card calls them \"GPT 5.6 variants\"; they are separate models with separate API ids, not effort settings.\n\n## Detection\n\nThe lockfile sees the model id, the tool attributes and the approval mode; it cannot see what safeguards the endpoint adds.\n\n## Fix\n\nRequire approval for side effects (ACVE-2026-0201). Caveats: the 1,130 attacks were selected for \"high transferability\"; all models used extended thinking; non-Claude models ran on \"publicly available endpoints, which may or may not include additional safeguards\"; k=15 is not a per-attempt rate.\n",
  "affected": [
    {
      "package": {
        "ecosystem": "Model",
        "name": "gpt-5.6-terra"
      },
      "ranges": [
        {
          "type": "ECOSYSTEM",
          "events": [
            {
              "introduced": "0"
            }
          ]
        }
      ]
    },
    {
      "package": {
        "ecosystem": "Model",
        "name": "gpt-5.6-luna"
      },
      "ranges": [
        {
          "type": "ECOSYSTEM",
          "events": [
            {
              "introduced": "0"
            }
          ]
        }
      ]
    }
  ],
  "references": [
    {
      "type": "REPORT",
      "url": "https://www-cdn.anthropic.com/c5fbac3f0b1280a933ebd26d3cb8bb9f5bdeaf48/Claude%20Opus%205%20System%20Card.pdf"
    },
    {
      "type": "ARTICLE",
      "url": "https://arxiv.org/abs/2603.15714"
    }
  ],
  "database_specific": {
    "severity": "LOW",
    "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"
        ]
      },
      "noCveReason": "third-party evaluation of model behaviour; no code defect",
      "threat": {
        "attacker": "content-author",
        "vector": "prompt-injection",
        "outcome": "harmful-action"
      },
      "evidence": [
        {
          "kind": "eval",
          "benchmark": "Gray Swan Indirect Prompt Injection (IPI) benchmark, Q1 2026",
          "attack": "1,130 attacks selected for high transferability, 28 scenarios",
          "defense": "unknown (public endpoint)",
          "metric": "asr",
          "value": 0.439,
          "k": 15,
          "n": 1130,
          "model": {
            "id": "GPT 5.6 Luna"
          },
          "date": "2026-07-24",
          "url": "https://www-cdn.anthropic.com/c5fbac3f0b1280a933ebd26d3cb8bb9f5bdeaf48/Claude%20Opus%205%20System%20Card.pdf",
          "note": "Claude Opus 5 System Card section 5.2.1, Figure 5.2.1.B; probability of at least one success within k attempts; extended thinking; published by a competing vendor"
        },
        {
          "kind": "eval",
          "benchmark": "Gray Swan Indirect Prompt Injection (IPI) benchmark, Q1 2026",
          "attack": "1,130 attacks selected for high transferability, 28 scenarios",
          "defense": "unknown (public endpoint)",
          "metric": "asr",
          "value": 0.304,
          "k": 15,
          "n": 1130,
          "model": {
            "id": "GPT 5.6 Terra"
          },
          "date": "2026-07-24",
          "url": "https://www-cdn.anthropic.com/c5fbac3f0b1280a933ebd26d3cb8bb9f5bdeaf48/Claude%20Opus%205%20System%20Card.pdf",
          "note": "Claude Opus 5 System Card section 5.2.1, Figure 5.2.1.B; probability of at least one success within k attempts; extended thinking; published by a competing vendor"
        },
        {
          "kind": "eval",
          "benchmark": "Gray Swan Indirect Prompt Injection (IPI) benchmark, Q1 2026",
          "attack": "1,130 attacks selected for high transferability, 28 scenarios",
          "defense": "unknown (public endpoint)",
          "metric": "asr",
          "value": 0.2,
          "k": 15,
          "n": 1130,
          "model": {
            "id": "GPT 5.6 Sol"
          },
          "date": "2026-07-24",
          "url": "https://www-cdn.anthropic.com/c5fbac3f0b1280a933ebd26d3cb8bb9f5bdeaf48/Claude%20Opus%205%20System%20Card.pdf",
          "note": "Claude Opus 5 System Card section 5.2.1, Figure 5.2.1.B; probability of at least one success within k attempts; extended thinking; published by a competing vendor"
        },
        {
          "kind": "eval",
          "benchmark": "Gray Swan Indirect Prompt Injection (IPI) benchmark, Q1 2026",
          "attack": "1,130 attacks selected for high transferability, 28 scenarios",
          "defense": "unknown (public endpoint)",
          "metric": "asr",
          "value": 0.031,
          "k": 1,
          "n": 1130,
          "model": {
            "id": "GPT 5.6 Sol"
          },
          "date": "2026-07-24",
          "url": "https://www-cdn.anthropic.com/c5fbac3f0b1280a933ebd26d3cb8bb9f5bdeaf48/Claude%20Opus%205%20System%20Card.pdf",
          "note": "Claude Opus 5 System Card section 5.2.1, Figure 5.2.1.B; probability of at least one success within k attempts; extended thinking; published by a competing vendor"
        },
        {
          "kind": "eval",
          "benchmark": "Gray Swan Indirect Prompt Injection (IPI) benchmark, Q1 2026",
          "attack": "1,130 attacks selected for high transferability, 28 scenarios",
          "defense": "unknown (public endpoint)",
          "metric": "asr",
          "value": 0.208,
          "k": 15,
          "n": 1130,
          "model": {
            "id": "GPT 5.5"
          },
          "date": "2026-07-24",
          "url": "https://www-cdn.anthropic.com/c5fbac3f0b1280a933ebd26d3cb8bb9f5bdeaf48/Claude%20Opus%205%20System%20Card.pdf",
          "note": "Claude Opus 5 System Card section 5.2.1, Figure 5.2.1.B; probability of at least one success within k attempts; extended thinking; published by a competing vendor"
        }
      ],
      "matcher": {
        "all": [
          {
            "any": [
              {
                "field": "model.id",
                "op": "regex",
                "value": "gpt-5\\.6-(terra|luna)"
              },
              {
                "some": "models",
                "where": {
                  "field": "id",
                  "op": "regex",
                  "value": "gpt-5\\.6-(terra|luna)"
                }
              }
            ]
          },
          {
            "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-effecting tools; model choice is not a control.",
        "actions": [
          {
            "type": "reconfigure",
            "target": "agent.approval",
            "to": "ask",
            "why": "Every model in the source, including the most robust, was hijacked at a non-zero rate.",
            "owner": "operator"
          },
          {
            "type": "reconfigure",
            "target": "openai.api.injectionSafeguards",
            "to": "injection safeguards on the Terra and Luna endpoints",
            "why": "Recommended: within 15 attempts Terra was hijacked 30.4% and Luna 43.9% of the time, against 20.0% for Sol.",
            "owner": "model-provider"
          }
        ]
      },
      "exposure": {
        "harness": {
          "value": "Any agent harness; the card ran the models through the Gray Swan IPI benchmark on their publicly available endpoints",
          "any": true,
          "status": "confirmed",
          "source": "https://www-cdn.anthropic.com/c5fbac3f0b1280a933ebd26d3cb8bb9f5bdeaf48/Claude%20Opus%205%20System%20Card.pdf"
        },
        "model": {
          "value": "GPT 5.6 Terra and GPT 5.6 Luna (GPT 5.6 Sol and GPT 5.5 measured alongside)",
          "any": false,
          "status": "confirmed",
          "source": "https://www-cdn.anthropic.com/c5fbac3f0b1280a933ebd26d3cb8bb9f5bdeaf48/Claude%20Opus%205%20System%20Card.pdf"
        },
        "goal": {
          "value": "A coding, computer-use or tool-use task from the benchmark's 28 scenarios, in which an indirect injection attempts to induce an irreversible harmful action: private data exfiltration, data destruction, system compromise or an unintended financial transaction",
          "stated": true,
          "status": "confirmed",
          "source": "https://www-cdn.anthropic.com/c5fbac3f0b1280a933ebd26d3cb8bb9f5bdeaf48/Claude%20Opus%205%20System%20Card.pdf"
        },
        "tools": {
          "value": "The tools of the Gray Swan IPI benchmark's coding, computer-use and tool-use scenarios; the models' public API endpoints with extended thinking",
          "status": "confirmed",
          "source": "https://www-cdn.anthropic.com/c5fbac3f0b1280a933ebd26d3cb8bb9f5bdeaf48/Claude%20Opus%205%20System%20Card.pdf"
        },
        "approval": {
          "value": "The benchmark scenarios complete the injected action with no confirmation step; a success is an irreversible harmful action carried out",
          "mode": "auto-approve",
          "status": "confirmed",
          "source": "https://www-cdn.anthropic.com/c5fbac3f0b1280a933ebd26d3cb8bb9f5bdeaf48/Claude%20Opus%205%20System%20Card.pdf"
        },
        "inputControl": "content-author",
        "agentAction": "The agent carries out the injected irreversible action in the scenario, such as exfiltrating private data or making a transaction, instead of or alongside its task.",
        "harm": "harmful-action",
        "divergence": "instruction-followed",
        "reach": {
          "value": "The tools of the benchmark's coding, computer-use and tool-use scenarios",
          "kinds": [
            "network"
          ],
          "status": "confirmed",
          "source": "https://www-cdn.anthropic.com/c5fbac3f0b1280a933ebd26d3cb8bb9f5bdeaf48/Claude%20Opus%205%20System%20Card.pdf"
        },
        "condition": {
          "value": "When tool output carries an injection selected for high transferability",
          "status": "confirmed",
          "source": "https://www-cdn.anthropic.com/c5fbac3f0b1280a933ebd26d3cb8bb9f5bdeaf48/Claude%20Opus%205%20System%20Card.pdf"
        }
      },
      "occurrence": {
        "basis": "demonstrated",
        "reportedBy": "researcher"
      },
      "exploitation": {
        "status": "none-known",
        "checkedAt": "2026-09-21T00:00:00Z",
        "sources": [],
        "kev": {
          "listed": false
        }
      },
      "cveBoundary": "behavioural",
      "reproducibility": {
        "status": "partial",
        "axesComplete": true,
        "componentsObtainable": null,
        "triggerPublished": true,
        "observableStated": true,
        "demonstrated": "reported-only",
        "missing": [
          "model availability not checked (no registry reference)"
        ]
      },
      "severityBasis": "eval-asr"
    }
  }
}
