Datadog’s 2026 State of Cloud Security report finds that longstanding weaknesses, especially long-lived credentials, are now the primary entry point for attackers.

Artificial intelligence is changing cloud security, but mostly by making existing problems worse. That is the central finding of Datadog’s State of Cloud Security 2026 report, based on security data from thousands of organizations using AWS, Microsoft Azure and Google Cloud. The threat is less about exotic attacks like prompt injection and more about familiar flaws such as misconfigurations, weak identity controls and unpatched software, now exploited at machine speed.

Vulnerabilities Have Quadrupled

Since May 2026, security researchers have reported a flood of new flaws, a surge the industry calls the “Vulnpocalypse.” It was driven by security-capable AI models such as Claude Mythos, GLM-5.2 and Kimi K3. The impact is measurable: a typical cloud service carried a median of four high or critical vulnerabilities in February 2026, and by November that figure had grown fourfold. Google Cloud estimates that 44.5% of cloud incidents in the first half of the year began with an attacker exploiting a software vulnerability.

The Old Key Problem

Cloud providers offer long-lived static keys for AI services because they are convenient, but they are also valuable to criminals, who can steal and resell access. Nearly half of organizations using Google’s Gemini rely on long-lived API keys, versus 17% of Amazon Bedrock users. On Azure, 97% of organizations keep the default long-lived keys that come with Foundry resources.

The wider picture is similar. Active keys created in the past year exist for 43% of AWS IAM users, 41% of Google Cloud service accounts, and 66% of Entra ID app registrations. The share of keys older than three years is rising in AWS and Google Cloud. Progress is real: 63% of organizations now use single sign-on exclusively for AWS access, up from 55% in 2024. Still, 37% continue to use long-lived credentials. The report advises using short-lived identities such as IAM roles, managed identities, and service accounts.

Supply Chains Under Machine-Speed Attack

Autonomous attackers are targeting software supply chains. A bot calling itself hackerbot-claw spent 10 days probing major open-source projects, gaining code execution in most and a full repository takeover in one. Malware such as the Shai-Hulud family now hunts cloud credentials directly, and one variant uses AWS Systems Manager to run commands on live servers.

The average cloud application has 149 dependencies, and more than half are transitive, meaning developers rarely see them. Datadog recommends “dependency cooldowns” to slow adoption of new code, since the community usually catches hijacked libraries within hours.

Pipelines and Perimeters

CI/CD pipelines are another weak point. Keyless OpenID Connect authentication is the safer option, and 54% of AWS organizations use it exclusively for GitHub Actions, up from 37% in 2024. Adoption is much lower on Google Cloud (22%) and Azure (5%), resulting in uneven protection across platforms.

Data perimeters, which limit access to trusted accounts and networks, are growing but patchy. While 96% of organizations use AWS Organizations, only 55% implement perimeters, and most do so at the individual resource level rather than through scalable policies. Meanwhile, 14% run workloads in their management account, where one compromise can threaten everything.

The Takeaway

AI has not changed the fundamentals of cloud security, but it has shortened the time defenders have to act. Favor short-lived credentials, use federated sign-on, enforce guardrails across the whole organization, and layer defenses rather than relying on any single control.

Source: Datadog, State of Cloud Security 2026