On November 10, 2025, cybersecurity firm Wiz published a blistering report revealing security gaps among 50 leading AI companies worldwide. The report shows that 65% of the companies scanned had exposed verified sensitive secrets in GitHub code repositories—including API keys, tokens and credentials—collectively representing firms with a total valuation of more than $400 billion. The finding reverberated through Silicon Valley and was seen as a wake‑up call against the AI sector's 'speed‑first' strategy. Wiz researchers noted that such lapses not only expose internal data but can also affect partner supply chains. Media outlets including TechRepublic and SecurityWeek quickly followed up, warning that AI security has become an 'invisible bomb' for the industry.
Wiz focused its analysis on the Forbes AI 50 list, applying a three‑dimensional scanning approach it calls 'depth, perimeter and coverage' to examine full commit histories, branches, deleted repositories, workflow logs and other 'hidden areas'. The results were striking: leaks were often concealed in places beyond the reach of conventional scanners, such as personal repositories or third‑party contributor code. Glyn Morgan, national manager for Salt Security in the UK and Ireland, told TechCrunch: 'Pushing credentials into code repositories is tantamount to handing attackers a golden key that bypasses all defenses.' OWASP classifies these issues as governance and configuration risks. Morgan added: 'These are basic, avoidable mistakes that get overlooked in fast‑paced iteration.'
The report lists multiple real‑world cases that underscore the severity of the problem. For example, LangChain leaked several Langsmith API keys, some of which had permissions to manage the organization and list members—intelligence highly valuable for attacker reconnaissance. Another Forbes AI 50 company left an ElevenLabs enterprise API key in plaintext. More severely, an unnamed AI 50 company exposed a HuggingFace token in a deleted branch that allowed access to roughly 1,000 private models; the same company also leaked a WeightsAndBiases key, revealing training data for several private models. Wiz researchers warn these exposures can disclose organizational structure, training datasets and even private models themselves—effectively handing over core AI assets. The combined valuation of affected companies tops $400 billion, far exceeding the typical scale of security incidents in conventional software firms.
The root cause, the report argues, is the AI industry's 'race culture.' To gain advantage, AI firms often adopt an agile 'code‑as‑infrastructure' development model where developers iterate rapidly on platforms like GitHub but neglect secret‑management best practices. Traditional security scans focus on core repositories, whereas Wiz's 'perimeter scan' extends to organization members and contributors' personal accounts, which can unintentionally commit corporate credentials to public repositories. Researchers identified these 'adjacent accounts' by tracing code contributors, organization followers and related networks such as HuggingFace and npm. Its 'coverage scan' aims at AI‑specific secret types (for example, WeightsAndBiases keys and Groq platform tokens) that many conventional tools miss. The report compares this to an iceberg: visible risks are only the tip, while lurking threats are far more dangerous.
The disclosure process also revealed immaturity in many AI companies. When Wiz attempted to report incidents to affected firms, nearly half of notifications were either undeliverable or went unanswered. Many companies lacked official disclosure channels or took no action after being notified. Forbes reported this as evidence of a 'security governance vacuum' among AI startups: although funding is high (global AI investment exceeded $500 billion in 2025), security budgets account for only 1–2%. Morgan warned: 'When enterprises collaborate with AI startups, they may inherit their security posture; one leak can create risk across the entire supply chain.'
Expert commentary amplified the alarm. The head of OWASP's API security project called these leaks typical governance‑and‑configuration risks, noting attackers can use exposed keys to access systems, data and models and thereby bypass firewalls. Morgan added: 'AI innovators must understand that speed cannot come at the expense of security.' In a Reuters interview, Wiz CEO Assaf Rappaport emphasized: 'We scanned the Forbes AI 50 and found 65% leakage—this is not an isolated incident but a systemic problem.' Gizmodo reported similar incidents were already frequent in 2024 (for example, OpenAI key leaks that led to model access abuse), and that incidents rose about 30% in 2025.
The global impact reaches into supply chains. Wiz warns organizations adopting tools from AI startups must scrutinize those vendors' secret‑management and vulnerability‑disclosure practices. The report offers three action recommendations: 1) treat employees as part of the attack surface and establish VCS member policies (for example, enforce MFA and separate multiple accounts); 2) expand secret scanning to the 'depth, perimeter, coverage' model, covering history, branches and personal repositories; 3) prioritize detection of AI‑specific secret types (such as HuggingFace tokens). Rappaport concluded: 'AI companies are valued at $400 billion, but security lapses can evaporate that value overnight.'
More than a warning, the report is a call to action. TechCrunch analysis predicts AI‑related security incidents could account for 30% of cyberattacks in 2025, and companies must shift from a 'speed race' to a 'security race.' OWASP plans to update API risk guidance to include AI‑specific vulnerabilities. A Forbes columnist wrote that Wiz's findings 'remind investors: beneath the AI bubble lie security landmines.' Amid a $500 billion global AI investment wave, this leak crisis highlights that innovation must be pursued alongside protection, or the 'golden age' could turn into a 'disaster age.'