Company AI Risk Explorer
Loading the data…
Welcome to the Company AI Risk Explorer

This tool maps what large global companies say in public about AI risk. We reviewed public documents from roughly 200 of the world's largest companies. We coded every risk they acknowledge against the AI Risk Repository taxonomy.

Filter by company, sector, AI role, or risk. Compare what companies acknowledge against expert ratings of severity, sector vulnerability, and role responsibility.

← Back to the overview Company AI Risk Explorer

Companies × risks

Meta

Meta develops general-purpose foundation models (the Llama family, including the 405-billion-parameter Llama 3.1 trained on over 16,000 H100 GPUs) released under open-source licences. It also builds specialised recommendation and ads ranking models at massive scale, and participates in AI governance through the Partnership on AI.

We reviewed 10 public documents from Meta and found it acknowledges 11 of 24 AI risks, including weapons & cyberattacks and AI security vulnerabilities.

Sector: Information

Roles: General-purpose AI developer; Specialized AI developer; AI deployer

AI risks Meta acknowledges

  • Weapons & cyberattacks: “Identifying catastrophic outcomes to prevent: Our framework identifies potential catastrophic outcomes related to cyber, chemical and biological risks that we strive to prevent.” (Our Approach to Frontier AI)
  • AI security vulnerabilities: “our efforts to develop and deploy AI models, internal and third-party AI tools, and other AI applications expose us to increased and novel risks and vulnerabilities, including prompt injection... as well as compromise of valuable intellectual property including source code, model weights, and other assets.” (Meta - Form 10-K 2025 Annual Report)
  • Toxic content: “we evaluated each fine-tuned model against a range of risks, including in areas like violent crime, child sexual exploitation” (Connect 2024: The Responsible Approach We're Taking to Generative AI)
  • Loss of privacy: “We employed different techniques to reduce memorization... we red-teamed the model for memorization of information about private individuals and took steps to mitigate the model against those risks.” (Expanding Our Open Source Large Language Models Responsibly)
  • Devaluation of human creativity: “litigation... alleging copyright infringement in connection with the acquisition, distribution, and use of copyrighted materials for AI training as well as potential reproduction of copyrighted materials in AI outputs” (Meta - Form 10-K 2025 Annual Report)
  • Capability & robustness: “it is possible that the actual numbers of unique people using our products may vary significantly from our estimates, potentially beyond our estimated error margins… our data limitations may affect our understanding of certain details of our business and increase the risk of error for our Family metrics estimates.” (Meta - Form 10-K 2025 Annual Report)
  • False information: “subject us to risks related to... accuracy, misinformation and deepfakes” (Meta - Form 10-K 2025 Annual Report)
  • Dangerous capabilities: “We define our thresholds based on the extent to which frontier AI would uniquely enable the execution of any of the threat scenarios we have identified as being potentially sufficient to produce a catastrophic outcome.” (Meta's Frontier AI Framework)
  • Fraud & scams: “propensity to automate social engineering via spear-phishing” (Expanding Our Open Source Large Language Models Responsibly)
  • Discrimination: “we cannot guarantee that third parties will not use such AI technologies for improper purposes, including... furthering bias or discrimination” (Meta - Form 10-K 2025 Annual Report)
  • Disinformation & influence: “...accuracy, misinformation and deepfakes (including related to elections) ...” (Meta - Form 10-K 2025 Annual Report)

Documents we reviewed