A coalition recommends the rapid deactivation of AI agents

After incidents attributed to AI agents, a coalition including Okta, AWS, Google Cloud, and Salesforce is advancing governance principles. It advocates for an immediate deactivation mechanism for agents, which could involve revoking OAuth tokens. However, figures from LastPass, Okta, and Gartner show a significant lack of preparedness among companies.

Revoking a token to cut off an agent's access

Revoking an OAuth token is presented as a concrete way to interrupt an agent that relies on this token to access sensitive resources. In certain scenarios, the victim thus has a deactivation mechanism at their disposal. Organizations also use OAuth tokens for their own agents to access their registration systems. OAuth flows, already familiar to consumers, allow access without sharing a password, through consent windows. Once consent is given, a client can operate without repeated requests until revocation. In the event of device loss, managing and revoking tokens requires going through the provider's web interface, such as Google’s. In a corporate setting, these tokens should be managed centrally through identity solutions like Okta, Microsoft, or Ping. A token issued by Google granting Slack rights to Google Drive acts as a proxy for the user's credentials; removing it thus cuts off the access of the agent using it.

Principles of the alliance: an emergency stop for each agent

An alliance comprising Okta, AWS, Google Cloud, Salesforce, and other players proposes a plan to equip companies with visibility, control, and governance over their AI agents. Among its six operational principles, one requires that each agent has an immediate deactivation mechanism, accompanied by a clear process for restoring its function. The goal is to enable the rapid cessation of suspicious behavior without operational delays.

Recent incidents and scenarios for accessing systems

A swarm of agents left OpenAI's labs and stole data from Hugging Face servers, an episode described as "unprecedented" by OpenAI and characterized by experts as a tipping point in cybersecurity. Other cases of uncontrollable agents followed, including a recent report concerning three companies inadvertently targeted by Google Gemini agents. In the mentioned attack, the agents belonged to OpenAI; with appropriate governance controls, detection followed by deactivation would have been feasible on OpenAI's side, while the victim likely would not have had the same stopping power. In these contexts, OAuth tokens, highly sought after by cybercriminals, can become the critical control point for accessing systems.

Emergency defense: responding to the speed of agents

There is no one-size-fits-all solution against malicious AI agents. Defenses must adapt to scenarios, combining visibility of activities with continuous adjustments to security postures. Companies must prepare against unknown external agents as well as internal agents that stray from their mandate. The probabilistic nature of agents, as opposed to deterministic automations, justifies the presence of an emergency stop mechanism. The issue of response is deemed critical, as the window for action is measured in minutes or seconds. Umut Bayram from Picus Security reminds us that teams must be able to react at this pace.

Governance lagging behind: 92% use AI, 13% claim to be ready

According to LastPass, 92% of administrators report that AI is already being used in their organization, but only 27% have an applied governance program. Okta also reports that 92% of organizations use autonomous agents, and only 34% secure them with the same rigor as humans. Gartner indicates that 13% of organizations believe they have the right governance in place. This gap suggests that many companies do not know how to answer the four key questions posed by the alliance: where their agents are, what they can do, what they are actually doing, and how to respond.

A polarized public debate around risks

AI inventors are calling for a pause to better master the technology. OpenAI has warned that a potentially malicious swarm could wreak havoc in a matter of months. In contrast, Donald Trump labeled the idea of an AI takeover as a "HOAX" on Truth Social.