In late January 2026, the social network Moltbook—designed to allow AI agents to post and interact autonomously— reported having between 1.4 and 1.5 million active accounts, before revealing a misconfigured database that exposed email addresses, login tokens, and API keys for all registered agents. The incident, linked to the autonomous assistant OpenClaw, illustrates a problem that most companies have not yet fully grasped: their AI agents, connectors, and integrations rely on identities (API keys, service accounts, OAuth tokens) that are managed with far less rigor than privileged human accounts.
Non-Human Identities (NHI): A Surge in Business Risks
According to an analysis by the Forum éWorld Economic Forum, non-human identities (NHIs)—such as service accounts, API keys, and authentication tokens—are rapidly proliferating in enterprise environments, creating new challenges for organizations in terms of visibility, governance, and cybersecurity.
The Semper Reporti2026, which surveyed IT and security professionals in eight countries, including France, quantifies the scale of the trend: nearly 29% of organizations are already using autonomous agents for privileged tasks such as resetting passwords or granting VPN access, and 65% plan to take this step by next year. This figure far exceeds the proportion of organizations that say they are prepared to handle the consequences of an incident involving these same agents.
Why Traditional IAM Fails in the Face of AI Agents’ Autonomy
Traditional IAM (Identity Access Management) was designed for humans and static applications, with monthly or quarterly access review cycles. An AI agent dynamically switches tasks, can invoke multiple tools in a single session, and sometimes operates at machine speed with no human intervention between the request and execution. This mismatch in pace explains the most widely documented vulnerability: many agents still use shared human credentials, generic API keys, or shared service accounts, due to the lack of an identity model designed specifically for them.
According to a Delinea study, 80% of companies report that they are not always able to understand why a non-human identity performed a privileged action—a lack of traceability that becomes a direct regulatory risk whenever an incident involves personal data or critical systems. The same study notes that 90% of organizations face internal pressure to relax their identity controls in order to accelerate AI projects, at the expense of governance.
5 Best Practices for Securing AI Access Management
- A unique identity for each agent: never shared human credentials or a generic service account shared among multiple agents, to enable precise attribution of each action.
- Minimal scope and explicit permissions: Each API key or token accesses only the resources necessary for the agent’s task, never a broad scope “just in case.”
- Automatic rotation: keys with a limited lifespan rather than static credentials that are valid indefinitely, using secret vaults (Vault, AWS Secrets Manager, etc.) rather than hard-coded keys in the code or configuration files.
- Immediate, centralized revocation: An up-to-date inventory of all active agents and connectors makes it possible to revoke access within minutes in the event that an agent is compromised or an employee responsible for an agent leaves the company.
- Delegation traceability: For every action taken by an agent, it must be possible to answer the question, “On whose behalf is the agent acting, and who is responsible for that action?” This requires systematic logging linked to the agent’s identity, not just that of the original user.
OpenClaw, OWASP, and API Key Leaks: The Most Common Vulnerabilities
The Moltbook and OpenClaw incidents, documented in detail in early 2026, combine two categories of failures that are worth distinguishing. On the one hand, there are classic development errors (API keys hard-coded in a public Git repository, a misconfigured proxy, a database exposed without authentication) that are not specific to AI and fall under basic application security best practices. On the other hand, there are genuinely new risks related to autonomy: poorly defined access delegation, agents capable of acting on multiple systems simultaneously if a single key is compromised, and exposure to prompt injection, which can divert a legitimate agent from its original objective. In fact, the OWASP framework dedicated to agent-based applications lists identity and privilege abuse among its most critical risks for 2026, alongside mission hijacking and tool misuse.
Cybercrime and AI tokens: a significantly expanded scope of exposure
The SpyCloud2026report on identity exposure documents a concrete criminal dimension: 18.1 million exposed API keys and tokens, in addition to 6.2 million authentication credentials and cookies specific to AI tools, observed circulating in sources originating from cybercriminal networks. When agent credentials are compromised, the impact far exceeds that of a typical human account: an agent often has cross-system access to multiple systems that no individual user would normally have, which multiplies the scope of exposure from a single data breach.
→ See our article on AI governance in the enterprise for a comprehensive overview of how it integrates with other controls.
FAQ: AI Identity and Access Security
Does an AI agent need to have its own account, separate from that of the employee who set it up?
Is a specific tool needed to manage the identities of AI agents?
How can we prioritize if we can't audit all machine identities at once?
Secure the identities of your agents and AI connectors
Test the security of your API keys, OAuth tokens, and service accounts under real-world attack conditions before a leak occurs. Would you like to discuss your project or assess your needs? Contact our experts.