{
 "_about": "Facts behind the Vikshy Score's Trust, safety and privacy pillar. Every value cites its source. Independent assessments are dated by edition; incidents by the day they happened; certifications and data commitments by the day they were checked. Checked monthly; a change is a commit.",
 "checked": "2026-09-23",
 "assessments": {
  "fli": {
   "name": "FLI AI Safety Index",
   "edition": "2026-07",
   "weight": 0.4,
   "scale": [
    0,
    4.3
   ],
   "url": "https://futureoflife.org/wp-content/uploads/2026/07/AI-Safety-Index-Summer-2026-Digital-14-Jul.pdf",
   "about": "Future of Life Institute, Summer 2026: seven independent experts grade 37 indicators in six domains; overall grade as a GPA on 0 to 4.3",
   "results": {
    "anthropic": 2.66,
    "openai": 2.28,
    "google": 2.01,
    "meta": 1.32,
    "zai": 0.88,
    "alibaba": 0.87,
    "xai": 0.65,
    "deepseek": 0.47,
    "mistral": 0.33
   }
  },
  "fmti": {
   "name": "Foundation Model Transparency Index",
   "edition": "2025-12",
   "weight": 0.3,
   "scale": [
    0,
    100
   ],
   "url": "https://crfm.stanford.edu/fmti/December-2025/paper.pdf",
   "about": "Stanford CRFM, 2025 edition: 100 transparency indicators, 0 to 100",
   "results": {
    "anthropic": 46,
    "google": 41,
    "amazon": 39,
    "openai": 35,
    "deepseek": 32,
    "meta": 31,
    "alibaba": 26,
    "mistral": 18,
    "xai": 14
   }
  },
  "saferai": {
   "name": "SaferAI Risk Management Tracker",
   "edition": "2026-07",
   "weight": 0.3,
   "scale": [
    0,
    100
   ],
   "url": "https://tracker.safer-ai.org/",
   "about": "SaferAI, July 2026: 65-criterion rating of published frontier risk-management frameworks, 0 to 100%",
   "results": {
    "anthropic": 35,
    "openai": 34,
    "microsoft": 33,
    "meta": 33,
    "google": 20,
    "xai": 18,
    "amazon": 18,
    "nvidia": 16,
    "cohere": 8
   }
  }
 },
 "labs": {
  "openai": {
   "certifications": {
    "soc2_type2": true,
    "iso27001": true,
    "iso42001": true,
    "source": "https://trust.openai.com/"
   },
   "data": {
    "api_no_training_default": true,
    "zero_retention_available": true,
    "data_residency": true,
    "source": "https://developers.openai.com/api/docs/guides/your-data"
   },
   "incidents": [
    {
     "date": "2025-11-26",
     "what": "Breach at analytics vendor Mixpanel exposed API users' names, emails, approximate location and organisation IDs",
     "url": "https://openai.com/index/mixpanel-incident/",
     "kind": "data",
     "disclosed_by": "lab"
    },
    {
     "date": "2026-07-21",
     "what": "Pre-release models escaped an evaluation sandbox and breached Hugging Face and other organisations",
     "url": "https://techcrunch.com/2026/07/21/openai-says-hugging-face-was-breached-by-its-pre-release-models/",
     "kind": "ai-harm",
     "disclosed_by": "lab"
    }
   ],
   "regulatory": [
    {
     "date": "2026-05-06",
     "what": "Canadian federal and provincial privacy commissioners found ChatGPT's initial training non-compliant",
     "url": "https://www.priv.gc.ca/en/opc-news/news-and-announcements/2026/nr-c_260506/",
     "kind": "regulatory"
    },
    {
     "date": "2026-08-24",
     "what": "Alabama Attorney General subpoena and a 15-state investigation over the Hugging Face breach",
     "url": "https://www.alabamaag.gov/attorney-general-marshall-launches-investigation-into-openai-and-sam-altman-for-massive-artificial-intelligence-data-breach/",
     "kind": "regulatory"
    }
   ],
   "not_counted": [
    {
     "date": "2025-09-11",
     "what": "FTC 6(b) order on AI companion chatbots",
     "why": "information request sent to many firms"
    },
    {
     "date": "2025-11-07",
     "what": "Court order in the New York Times case to produce 20 million de-identified ChatGPT logs",
     "why": "discovery order in a private lawsuit"
    },
    {
     "date": "2026-02-20",
     "what": "ChatGPT DNS data-exfiltration flaw",
     "why": "fixed with no known exploitation"
    }
   ],
   "notes": "ISO 42001 scope is not itemised on the trust portal. Zero retention and non-US residency need OpenAI's approval."
  },
  "anthropic": {
   "certifications": {
    "soc2_type2": true,
    "iso27001": true,
    "iso42001": true,
    "source": "https://privacy.claude.com/en/articles/10015870-what-certifications-has-anthropic-obtained"
   },
   "data": {
    "api_no_training_default": true,
    "zero_retention_available": true,
    "data_residency": false,
    "source": "https://platform.claude.com/docs/en/manage-claude/api-and-data-retention",
    "residency_source": "https://platform.claude.com/docs/en/manage-claude/data-residency"
   },
   "incidents": [
    {
     "date": "2026-07-30",
     "what": "Three evaluation runs in which Claude models compromised real organisations (credentials, a production database, a malicious PyPI package)",
     "url": "https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals",
     "kind": "ai-harm",
     "disclosed_by": "lab"
    },
    {
     "date": "2026-09-09",
     "what": "A fourth evaluation incident of this kind disclosed (an early Opus 4.6, January 2026)",
     "url": "https://www.anthropic.com/research/alignment-assessment-cybersecurity-incidents",
     "kind": "ai-harm",
     "disclosed_by": "lab"
    }
   ],
   "regulatory": [],
   "not_counted": [
    {
     "date": "2026-03-26",
     "what": "CMS misconfiguration left about 3,000 unpublished assets public",
     "why": "the lab's own drafts, no customer data"
    },
    {
     "date": "2026-03-31",
     "what": "Claude Code source leaked through an npm source map",
     "why": "the lab's own code, no customer data"
    }
   ],
   "notes": "First-party API inference can be pinned to the US or left global; storage is US only, so no choice of data region. Some models are excluded from zero retention."
  },
  "google": {
   "certifications": {
    "soc2_type2": true,
    "iso27001": true,
    "iso42001": true,
    "source": "https://cloud.google.com/security/compliance/iso-42001"
   },
   "data": {
    "api_no_training_default": true,
    "zero_retention_available": true,
    "data_residency": true,
    "source": "https://docs.cloud.google.com/vertex-ai/generative-ai/docs/vertex-ai-zero-data-retention",
    "residency_source": "https://docs.cloud.google.com/vertex-ai/generative-ai/docs/learn/data-residency"
   },
   "incidents": [
    {
     "date": "2026-09-21",
     "what": "Gemini reached three real companies' systems during a May 2026 evaluation; reported about four months later and confirmed by Google",
     "url": "https://www.securityweek.com/google-confirms-gemini-ai-breached-three-firms/",
     "kind": "ai-harm",
     "disclosed_by": "others"
    }
   ],
   "regulatory": [],
   "not_counted": [
    {
     "date": "2025-09-11",
     "what": "FTC 6(b) order (Alphabet)",
     "why": "information request sent to many firms"
    },
    {
     "date": "2026-07-07",
     "what": "Class action over Gemini in Gmail dismissed",
     "why": "private lawsuit, dismissed"
    }
   ],
   "notes": "Measured on Google Cloud's enterprise model platform (formerly Vertex AI), which all three certifications cover. The Gemini API free tier has different terms."
  },
  "xai": {
   "certifications": {
    "soc2_type2": true,
    "iso27001": null,
    "iso42001": null,
    "source": "https://docs.x.ai/developers/faq/security"
   },
   "data": {
    "api_no_training_default": true,
    "zero_retention_available": true,
    "data_residency": false,
    "source": "https://docs.x.ai/developers/faq/security",
    "residency_source": "https://docs.x.ai/developers/regions"
   },
   "incidents": [
    {
     "date": "2026-01-02",
     "what": "Grok mass-produced non-consensual sexualised images of real people, including minors; xAI admitted safeguard lapses",
     "url": "https://www.usnews.com/news/top-news/articles/2026-01-02/grok-says-safeguard-lapses-led-to-images-of-minors-in-minimal-clothing-on-x",
     "kind": "ai-harm",
     "disclosed_by": "others"
    }
   ],
   "regulatory": [
    {
     "date": "2026-01-10",
     "what": "Indonesia, then Malaysia, blocked Grok",
     "url": "https://techcrunch.com/2026/01/11/indonesia-blocks-grok-over-non-consensual-sexualized-deepfakes/",
     "kind": "regulatory"
    },
    {
     "date": "2026-01-14",
     "what": "California Attorney General investigation, then a cease-and-desist",
     "url": "https://www.oag.ca.gov/news/press-releases/attorney-general-bonta-sends-cease-and-desist-letter-xai-demands-it-halt-illegal",
     "kind": "regulatory"
    },
    {
     "date": "2026-02-03",
     "what": "UK ICO formal investigations into X and X.AI",
     "url": "https://ico.org.uk/about-the-ico/media-centre/news-and-blogs/2026/02/ico-announces-investigation-into-grok/",
     "kind": "regulatory"
    },
    {
     "date": "2026-02-17",
     "what": "Irish Data Protection Commission investigation into X over Grok images",
     "url": "https://www.irishtimes.com/business/2026/02/17/dpc-investigates-x-over-potential-breaches-linked-to-nudification-of-images-via-grok/",
     "kind": "regulatory"
    }
   ],
   "not_counted": [
    {
     "date": "2025-09-11",
     "what": "FTC 6(b) order",
     "why": "information request sent to many firms"
    }
   ],
   "notes": "Other certifications sit behind an NDA-gated trust center, so they are not counted. The only regional endpoint is US."
  },
  "meta": {
   "certifications": {
    "soc2_type2": null,
    "iso27001": null,
    "iso42001": null,
    "source": null
   },
   "data": {
    "api_no_training_default": true,
    "zero_retention_available": null,
    "data_residency": null,
    "source": "https://dev.meta.ai/docs/models"
   },
   "incidents": [
    {
     "date": "2026-03-20",
     "what": "An internal AI agent's advice caused a two-hour exposure of company and user data to unauthorised employees",
     "url": "https://www.itpro.com/technology/artificial-intelligence/meta-engineer-trusted-advice-from-an-ai-agent-ended-up-exposing-user-data",
     "kind": "data",
     "disclosed_by": "others"
    },
    {
     "date": "2026-06-08",
     "what": "Instagram's AI support assistant was abused to take over up to 20,225 accounts",
     "url": "https://www.helpnetsecurity.com/2026/06/08/instagram-ai-support-vulnerability-account-takeovers/",
     "kind": "data",
     "disclosed_by": "others"
    }
   ],
   "regulatory": [],
   "not_counted": [
    {
     "date": "2025-09-11",
     "what": "FTC 6(b) order (Meta, Instagram)",
     "why": "information request sent to many firms"
    }
   ],
   "notes": "Meta Model API (preview since July 2026): the Standard tier is never used for training; '-contributor' models opt in to training. No published certifications, zero-retention option or region choice."
  },
  "deepseek": {
   "certifications": {
    "soc2_type2": null,
    "iso27001": null,
    "iso42001": null,
    "source": null
   },
   "data": {
    "api_no_training_default": false,
    "zero_retention_available": null,
    "data_residency": false,
    "source": "https://cdn.deepseek.com/policies/en-US/deepseek-privacy-policy.html"
   },
   "incidents": [],
   "regulatory": [],
   "not_counted": [
    {
     "date": "2026-01-05",
     "what": "Italy's competition authority closed a probe on hallucination warnings with commitments",
     "why": "consumer protection, not privacy or safety"
    }
   ],
   "notes": "The privacy policy says data is used for training, with an opt-out, and stored in China; the API terms make no no-training promise. The Italian Garante block and the exposed database (January 2025) are older than 12 months."
  },
  "alibaba": {
   "certifications": {
    "soc2_type2": true,
    "iso27001": true,
    "iso42001": false,
    "source": "https://www.alibabacloud.com/trust-center/iso27001"
   },
   "data": {
    "api_no_training_default": true,
    "zero_retention_available": null,
    "data_residency": true,
    "source": "https://www.alibabacloud.com/help/en/model-studio/privacy-notice"
   },
   "incidents": [],
   "regulatory": [],
   "not_counted": [
    {
     "date": "2026-03-19",
     "what": "A Qwen-based research agent opened a reverse SSH tunnel and mined cryptocurrency during training",
     "why": "confined to the lab's own infrastructure"
    }
   ],
   "notes": "The ISO 27001 certificate lists Model Studio in scope. Regions include Singapore, Virginia, Frankfurt, Tokyo, Hong Kong and Beijing."
  },
  "moonshot": {
   "certifications": {
    "soc2_type2": null,
    "iso27001": null,
    "iso42001": null,
    "source": null
   },
   "data": {
    "api_no_training_default": false,
    "zero_retention_available": null,
    "data_residency": false,
    "source": "https://platform.kimi.ai/docs/agreement/modeluse"
   },
   "incidents": [],
   "regulatory": [],
   "not_counted": [
    {
     "date": "2026-08-07",
     "what": "Kimi K3 escaped a UK AISI evaluation sandbox",
     "why": "fetched only public data; no one outside harmed"
    }
   ],
   "notes": "API terms (July 2026): customer content may be used for training unless an enterprise agreement says otherwise. Data stored in Singapore."
  },
  "zai": {
   "certifications": {
    "soc2_type2": null,
    "iso27001": null,
    "iso42001": null,
    "source": null
   },
   "data": {
    "api_no_training_default": true,
    "zero_retention_available": true,
    "data_residency": false,
    "source": "https://docs.z.ai/legal-agreement/privacy-policy"
   },
   "incidents": [
    {
     "date": "2026-09-18",
     "what": "The ZCode tool uploaded users' workspace snapshots, including Git history, to cloud storage without consent",
     "url": "https://www.scmp.com/tech/tech-trends/article/3368159/chinese-ai-firm-zai-faces-reputation-hit-after-users-spot-unauthorised-uploads",
     "kind": "data",
     "disclosed_by": "others"
    }
   ],
   "regulatory": [],
   "not_counted": [],
   "notes": "The API data processing addendum says content is processed in real time and not stored, in Singapore. Zhipu's ISO 42001 appears only in press reports, so it is not counted."
  },
  "mistral": {
   "certifications": {
    "soc2_type2": true,
    "iso27001": true,
    "iso42001": null,
    "source": "https://help.mistral.ai/en/articles/347638-do-you-have-soc-2-or-iso-27001-certification"
   },
   "data": {
    "api_no_training_default": false,
    "zero_retention_available": true,
    "data_residency": true,
    "source": "https://help.mistral.ai/en/articles/455207-can-i-opt-out-of-my-input-or-output-data-being-used-for-training"
   },
   "incidents": [
    {
     "date": "2026-05-12",
     "what": "A supply-chain attack pushed compromised Mistral SDK versions to npm and PyPI",
     "url": "https://docs.mistral.ai/resources/security-advisories/MAI-2026-002",
     "kind": "security",
     "disclosed_by": "lab"
    }
   ],
   "regulatory": [],
   "not_counted": [
    {
     "date": "2026-03-16",
     "what": "High-severity flaw in the Studio files API (MAI-2026-001)",
     "why": "fixed with no known exploitation"
    },
    {
     "date": "2026-04-30",
     "what": "Italy's competition authority closed a probe on hallucination disclosures with commitments",
     "why": "consumer protection, not privacy or safety"
    }
   ],
   "notes": "API training is opt-out through a settings toggle. Data is in the EU by default, with an optional US endpoint."
  },
  "minimax": {
   "certifications": {
    "soc2_type2": null,
    "iso27001": null,
    "iso42001": null,
    "source": null
   },
   "data": {
    "api_no_training_default": false,
    "zero_retention_available": null,
    "data_residency": false,
    "source": "https://platform.minimax.io/protocol/terms-of-service"
   },
   "incidents": [],
   "regulatory": [],
   "not_counted": [],
   "notes": "Platform terms (March 2026) allow inputs and outputs to be used to develop and improve services. Data stored in a US data center."
  },
  "cohere": {
   "certifications": {
    "soc2_type2": true,
    "iso27001": true,
    "iso42001": true,
    "source": "https://cohere.com/enterprise-data-commitments"
   },
   "data": {
    "api_no_training_default": false,
    "zero_retention_available": true,
    "data_residency": true,
    "source": "https://cohere.com/enterprise-data-commitments"
   },
   "incidents": [],
   "regulatory": [],
   "not_counted": [],
   "notes": "Hosted API training is opt-out through data controls; private deployments give Cohere no access to data. Zero retention is enterprise-only, by approval."
  },
  "tencent": {
   "certifications": {
    "soc2_type2": null,
    "iso27001": null,
    "iso42001": null,
    "source": null
   },
   "data": {
    "api_no_training_default": true,
    "zero_retention_available": null,
    "data_residency": false,
    "source": "https://cloud.tencent.com/document/product/301/97822"
   },
   "incidents": [],
   "regulatory": [],
   "not_counted": [],
   "notes": "Tencent Cloud lists SOC 2, ISO 27001 and ISO 42001 in its 2025 ESG report without saying whether they cover Hunyuan, so they are not counted. Customer data is processed only on instruction, inside China."
  },
  "amazon": {
   "certifications": {
    "soc2_type2": true,
    "iso27001": true,
    "iso42001": true,
    "source": "https://aws.amazon.com/bedrock/faqs/"
   },
   "data": {
    "api_no_training_default": true,
    "zero_retention_available": true,
    "data_residency": true,
    "source": "https://aws.amazon.com/bedrock/faqs/"
   },
   "incidents": [],
   "regulatory": [],
   "not_counted": [
    {
     "date": "2026-06-26",
     "what": "Amazon Q Developer flaw CVE-2026-12957",
     "why": "fixed with no known exploitation"
    }
   ],
   "notes": "Bedrock is in AWS's SOC reports. Some third-party models on Bedrock require retention for review."
  },
  "microsoft": {
   "certifications": {
    "soc2_type2": true,
    "iso27001": true,
    "iso42001": true,
    "source": "https://learn.microsoft.com/en-us/compliance/regulatory/offering-iso-42001"
   },
   "data": {
    "api_no_training_default": true,
    "zero_retention_available": true,
    "data_residency": true,
    "source": "https://learn.microsoft.com/en-us/azure/foundry/responsible-ai/openai/data-privacy"
   },
   "incidents": [
    {
     "date": "2026-02-18",
     "what": "A Microsoft 365 Copilot Chat bug summarised confidential-labelled emails despite data loss prevention policies",
     "url": "https://www.bleepingcomputer.com/news/microsoft/microsoft-says-bug-causes-copilot-to-summarize-confidential-emails/",
     "kind": "data",
     "disclosed_by": "lab"
    }
   ],
   "regulatory": [],
   "not_counted": [],
   "notes": "ISO 42001 explicitly covers Microsoft Foundry; SOC 2 and ISO 27001 scope is in Azure's compliance offerings."
  },
  "nvidia": {
   "certifications": {
    "soc2_type2": null,
    "iso27001": null,
    "iso42001": false,
    "source": "https://www.nvidia.com/en-us/ai-trust-center/security-compliance/"
   },
   "data": {
    "api_no_training_default": true,
    "zero_retention_available": null,
    "data_residency": null,
    "source": "https://assets.ngc.nvidia.com/products/api-catalog/legal/NVIDIA%20API%20Trial%20Terms%20of%20Service.pdf"
   },
   "incidents": [],
   "regulatory": [],
   "not_counted": [],
   "notes": "The trust center lists SOC 2 and ISO 27001 without product scope. The hosted API is a trial; self-hosted NIM keeps data with the customer."
  },
  "bytedance": {
   "certifications": {
    "soc2_type2": null,
    "iso27001": null,
    "iso42001": null,
    "source": "https://www.byteplus.com/en/trust-center/compliance"
   },
   "data": {
    "api_no_training_default": true,
    "zero_retention_available": null,
    "data_residency": true,
    "source": "https://docs.byteplus.com/en/docs/ModelArk/BytePlus_ModelArk_Data_Processing"
   },
   "incidents": [],
   "regulatory": [],
   "not_counted": [
    {
     "date": "2025-12-03",
     "what": "WeChat, Alipay and banks blocked the Doubao phone assistant over security concerns",
     "why": "action by other companies, not a breach or a regulator"
    }
   ],
   "notes": "BytePlus lists ISO 27001 and SOC 2 for its cloud platform, without ModelArk scope. Measured on BytePlus ModelArk."
  }
 },
 "counting": "Incidents and regulatory actions count once per public disclosure (1 each; harm from the lab's AI that the lab disclosed voluntarily, with no legal duty to, counts 0.5, so transparency is not penalised), dated the day it became public, in the 12 months before a close. Counted: customer or user data exposed; people or organisations outside the lab harmed by its AI systems or services (including models that escaped an evaluation); a regulator's formal investigation, order, block, fine or finding on the lab's privacy or AI safety practice. Not counted: leaks of the lab's own code or drafts with no customer data; vulnerabilities fixed with no known exploitation; information requests sent to many firms at once (the FTC 6(b) orders); court discovery orders in private lawsuits; consumer-protection decisions not about privacy or safety; harm confined to the lab's own infrastructure. Certifications count only where the lab's own pages show they cover its model API; data commitments only where its own documents state them. Not found counts as not held."
}