Cybersecurity: Google Prepares for Regulation of AI Agents

Google envisions a cybersecurity landscape by 2027 where specialized agent teams would take on part of the defense work. However, this scenario assumes controls over their access, actions, and deployment: without precautions, automation could also multiply the possibilities for attack.
Companies Must Audit Their Autonomous Workflows
Software vulnerabilities, faulty configurations, bypasses, indirect prompt injections: the integration of agents can expose companies to several categories of risks. The race for an AI-driven corporate image can push security checks to the background when adding features to applications or services. The attack surface could then expand beyond what was anticipated, even when AI is adopted to defend systems.
Auditing infrastructures and networks thus becomes a central precaution against adversaries who also possess AI, ranging from inexperienced hackers to state-sponsored groups. Penetration testing and offensive simulations by red teams should include agents, which were still outside the scope of many teams just a few years ago. Companies that have adopted the principle of integrating AI into their products and processes should identify and rectify corresponding vulnerabilities before an attacker can exploit them.
The recommendation for security leaders is to have autonomous workflows examined before any production rollout. The results of these tests should then be compared with threat intelligence. This combination should serve to strengthen network protections and harden classifiers against new attacks.
Google Aims to Contain the Proliferation of Agents
Google sees the proliferation of agents as a future form of ghost IT, with the accompanying risks. Often intended to operate alone under human control, these systems can become high-value access points for hackers when thousands of them operate in different environments. By 2027, the company envisions coordinated collectives capable of quickly handling numerous defense tasks.
Isolated agents equipped with general guidelines are deemed insufficient against AI-assisted adversaries, who can more rapidly compromise endpoints, infiltrate a network, navigate within it, or initiate data theft. The proposed model relies on orchestrated specialists: each should possess their training, domain knowledge, memory, and tools, with bounded autonomy. A central coordinator should determine when to solicit a human decision. This organization would require oversight to contain the uncontrolled proliferation of agents.
Their governance also involves an individual identity, akin to that of an employee, to integrate them into a zero-trust architecture and limit compromises. Heather Adkins, Vice President of Security Engineering at Google, recommends applying this logic to the transactions themselves, beyond the network. Even under zero trust, an agent could retain excessive freedom: thus, every transaction and authorization should be verified. Assigned rights should be minimal and revocable after an incident or when the agent is no longer useful.
Le brief IA que les pros lisent chaque soir
Les 7 actus IA du jour, décryptées en 5 min. Gratuit.
Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.
Choisissez votre rythme
Gratuit · Pas de spam · Désabonnement en 1 clic
Maintaining a clean history for each agent is also recommended, without manual management being feasible. Google envisions an authorization adjusted by real-time autonomous reasoning. An operation accepted on a Tuesday could thus be denied on Wednesday if its context appears abnormal.
Google Recommends Linking Production to Code
Software protection must also keep pace with the acceleration of its writing and deployment by AI. Google anticipates that by 2027, controls will need to be continuous within the delivery pipeline. Patches waiting several days and security checks reserved for the end of development are deemed inadequate: programming and securing should advance together, with agents, to maintain security despite the pace.
The threat to components is not new, as demonstrated by the SolarWinds attack in 2020. Sandra Joyce, Vice President of Google Threat Intelligence, states that cybercriminals now contaminate both libraries and open-source code. Programming assistants often automatically incorporate these components into codebases.
Google recommends directly connecting production environments to software sources. Defenders could then link a detected exposure to the relevant line and verify if it is indeed exploitable. The proposed goal is to provide a tested patch at an automated pace, without halting production.
Google Foresees Pre-Processed Alerts for Analysts
In Security Operations Centers (SOCs), the repetitive sorting of alerts occupies specialists, while many alerts are unnecessary or correspond to false positives. Teams are already familiar with the fatigue associated with this workload. In the face of autonomous threats capable of recognizing a network, penetrating it, and navigating within it in a matter of minutes, consulting traces scattered across multiple systems and days or weeks old does not allow for keeping pace.
Google anticipates the disappearance of the queue of pending alerts in its Cloud CISO Perspectives for 2027. Ramya Chitrakar, Vice President of Engineering at Google Cloud Security, believes that analysts will then be able to focus on directing results rather than pursuing alerts. The proposed change is to generate the alert after automated work, rather than making it the starting point for human intervention. Agents would enrich the data, correlate it, and perform triage to deliver a verified conclusion rather than an unaddressed ticket.
Specialists could monitor this processing and control automated interventions, allowing more availability for other tasks. Speed and processing volume are already the main motivations for companies considering AI for their security. According to security experts, agent capabilities are rapidly surpassing those of traditional SOCs in these two dimensions, while both defenders and attackers employ the technology.
Earlier this year, few organizations were using language models and agents in their cybersecurity operations, particularly to identify and rectify vulnerabilities. Google asserts that these uses are now being deployed massively. The company anticipates a large-scale influence of agents on security by 2027.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.