The median count of high or critical security flaws in cloud workloads jumped fourfold between May and November this year, according to Datadog's State of Cloud Security 2026 published November 10. The cloud observability and security firm attributes the spike to the influence of frontier AI systems and analyzed the security posture of thousands of customer organizations worldwide. The surge reflects vulnerabilities in real-world applications running in cloud environments, not just theoretical database entries, the company said.

Datadog reported that the median number of high or critical severity vulnerabilities climbed from four in May—when Anthropic released its Mythos model—to 14 in June and reached 16 by November. The data came from examining flaws in third-party libraries used by applications tracked through Datadog Code Security's Software Composition Analysis feature. Long-lived credentials present another major challenge, particularly for cloud-hosted AI inference services and CI/CD pipelines, with nearly half of organizations using Gemini relying on long-lived API keys. The share of cloud users maintaining access keys older than three years has hit nearly 30% for both AWS IAM and Google Cloud Service Account users.

The report warns that automated tools continuously scan for credential weaknesses, with red-team frameworks like Pacu and Datadog's own Pathrunner, alongside secret scanners like TruffleHog, reliably discovering over-permissioned credentials and exposed keys at scale. "That same tooling is available to autonomous attackers, who enumerate for long-lived credentials first because they're the highest-value, longest-lived prize in the environment," the report states. Rather than static API keys, organizations should access AI inference services through the calling workload's short-lived identity—such as IAM roles, managed identities, or service accounts—so models are reached with the same federated credentials as other cloud resources, Datadog recommends.

Supply-chain attacks are increasingly executed autonomously and at machine speed, with hijacked dependencies serving as one of the most popular compromise routes, according to the report. The danger stems partly from how quickly users adopt new dependency versions: 60% of organizations analyzed run versions released less than 24 hours earlier, with these versions appearing in more than a fifth of cloud applications. App workloads running in the cloud carry an average of 149 software dependencies, while the average organization manages 568. Datadog recommends implementing dependency cooldowns to reduce the risk of malicious dependencies reaching cloud environments, since the community typically identifies and contains hijacked libraries within hours. The escalating vulnerability count reinforces the need for cloud security foundations that limit the blast radius when applications are compromised. Organizations that fail to balance the velocity of AI-driven development with the discipline of credential hygiene and dependency vetting may discover that competitive advantage evaporates the moment an attacker exploits the same automation they rely on for innovation.