AI system security vulnerabilities and attacks

17 of the world's largest companies we reviewed acknowledge the AI risk "AI security vulnerabilities", including Microsoft, Alphabet (Google) and IBM.

Which companies acknowledge this risk?

• Microsoft: “protect against indirect prompt injection attacks; ... an attack mechanism where threat actors embed hidden malicious instructions in a grounding data source to circumvent safety guardrails” (Microsoft: G7 Hiroshima AI Process (HAIP) Transparency Report)
• Alphabet (Google): “SAIF is designed to help mitigate risks specific to AI systems like stealing the model, data poisoning of the training data, injecting malicious inputs through prompt injection” (Introducing Google's Secure AI Framework)
• IBM: “Jail breaking: A jailbreaking attack attempts to break through the guardrails established in the model to perform restricted actions... Prompt injection attacks... produce unexpected output by manipulating the structure, instructions, or information contained in its prompt.” (Foundation models: Opportunities, risks and mitigations)
• Amazon: “LLM-PIEval, a benchmark containing indirect prompt-injection attacks for LLMs... Attacks targeting sensitive APIs (e.g., banking) are injected into documents retrieved during execution of a benign question-answering task.” (Amazon Nova and our commitment to responsible AI)
• Salesforce: “Data poisoning: Attackers can change how an AI responds by injecting malicious or misleading content into the training data or documents used in Retrieval-Augmented Generation (RAG).” (Salesforce Trusted AI Impact Report)
• Meta: “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)
• SAP: “These attacks can be conducted directly or by placing adversarial prompts in web page content scraped by an autonomous agent.” (SAP Generative AI Cybersecurity Strategy)
• Tencent: “prevent jailbreak attacks and injection attacks while ensuring the performance of large-scale models are all urgent problems to be solved” (Large Model Safety and Ethics Research Report 2024 (announcement))
• GE Vernova: “Third-Party must maintain a catalog of trusted and valid data sources which will help protect against potential data poisoning or backdoor attacks.” (GE Vernova Third-Party Cyber Security Requirements)
• Cisco: “AI systems must be resilient and contain protections from malicious actors... Protection against security threats includes testing the resilience of AI systems against cyber-attacks; sharing information about vulnerabilities and cyber-attacks” (Cisco Principles for Responsible Artificial Intelligence)
• NVIDIA: “used open-source datasets to train our models... may otherwise decrease resilience to security incidents that may compromise the integrity of our AI outputs” (NVIDIA 2025 Annual Report and 10-K)
• Alibaba: “blue teaming... incorporates diverse perspectives to reveal potential or unknown risks... red team members... are encouraged to discover vulnerabilities... prompts for basic risks and attacks... emerging safety threats” (Alibaba Group 2025 ESG report)
• Mastercard: “These kinds of errors can be the result of either flawed logic or "data poisoning": bad actors' deliberate exposure of misleading data to the model.” (Mastercard Generative AI Cybersecurity Report)
• ServiceNow: “This reduces the introduction of irrelevant or low-quality information (as well as risks associated with "data poisoning")” (Responsible AI | ServiceNow)
• Commonwealth Bank of Australia: “CBA has implemented 11 guardrails to protect its models from risks related to malicious inputs” (Customer safety, convenience and recognition boosted by early ...)
• Newmont: “unauthorized access or exploitation of AI-powered systems, potentially compromising operations or sensitive data” (Annual Report 2025)
• Blackstone Group: “The volume and reliance on data and algorithms also make AI Technologies, and in turn us and our portfolio companies and investments, more susceptible to cybersecurity threats, including the compromise of underlying models, training data, or other intellectual property.” (2024 Blackstone 10-K filing)