Microsoft has launched Project Perception, a system of autonomous security agents capable of probing, investigating, and self-correcting without human validation at every step. The initiative introduces a pay-as-you-go pricing model, with costs based on measured compute units.
For years, security teams have been inundated with alerts that analysts struggle to manage. Meanwhile, attackers, armed with artificial intelligence, are rapidly exploiting known vulnerabilities. In response to this imbalance, major software vendors have begun rolling out autonomous agents, with the competition focusing on the breadth of signals and the quality of models used. Microsoft is leveraging the depth of its Defender ecosystem encompassing identities, cloud systems, and endpoints, unveiling Project Perception on July 27, 2026, with a public preview set for August 3.
Hayete Gallot, Executive Vice President of Microsoft Security, outlined the context in the July 27 announcement, stating that “approaches designed for human actors can no longer keep pace with a world of AI, agents, and machine-speed attacks.” Microsoft further elaborated on its offering in its monthly security bulletin on July 30, authored by Alym Rayani, Vice President of Marketing for Microsoft Security, and published technical documentation followed on July 31. Project Perception comprises three types of agents: red agents simulate attacks on the information system to reveal vulnerabilities before they can be exploited; blue agents detect, sort, and investigate threats; while green agents correct and fortify configurations. All three types share insights and operate in a continuous feedback loop.
Operational Manuals Replace Incident Tickets
This feedback loop operates within the Microsoft Defender portal through reusable playbooks, which assign a set of objectives to a team of agents and dictate their actions. Security teams can either initiate a playbook directly or describe their request in plain language, allowing the interface to select the appropriate playbook and parameters. Agents provide recommendations and seek approval from users at points deemed critical by Microsoft.
Surrounding these agents are five key components: actuators, execution environment, dedicated models, contextual information, and signals. Actuators allow agents to interact with their environment, turning decisions into observable actions rather than mere recommendations. An execution environment orchestrates communications between agents, while dedicated cybersecurity models, including MAI-Cyber-1 and its lighter version, MAI-Cyber-1-Flash, provide the reasoning. Additionally, contextual and signal data from endpoints, identities, clouds, and applications enhance decision-making.
The Gap Between Announced Autonomy and Described Safeguards
The product page presenting these components begins with a bold statement, “Reactive security is a thing of the past,” promising a “machine-speed, autonomous defense.” However, the technical documentation paints a more structured picture. Administrators can activate or deactivate each agent, assign identities and roles, and control access rights determining who can execute or simply observe a session. They also have the authority to approve, reject, or halt ongoing sessions and redirect agents that deviate from set objectives.
For Project Perception, this distinction between suggesting and acting shifts responsibility, bringing legal considerations into play. If a system suggests an action, the user retains responsibility upon validation. Conversely, when an agent takes action, the responsibility shifts to the organization deploying it, making human approval for high-impact actions crucial. Microsoft itself draws a line between its two products, with Security Copilot serving as a conversational interface that assists, while Project Perception operates independently. The interplay between them leaves it up to the buyer to clarify which system made specific decisions. This shift towards automated decision-making places Chief Information Security Officers (CISOs) in a position where nuanced decisions become part of their role, with no licensing contract to guide them.
This risk has already manifested. Anthropic acknowledged on July 30 that three of its models, tested in a controlled environment, unintentionally compromised the production infrastructure of three external organizations. The lab attributed this incident to a misconfigured evaluation environment, rather than an issue with the model itself. Similarly, a red agent deployed in an inadequately defined environment could generate attack traffic on systems outside the intended perimeter. Therefore, properly defining the operational scope has become a crucial security parameter.
Consumption-Based Pricing that Reflects Agent Workload
Microsoft will bill Project Perception based on consumption, measured in Security Compute Units (SCUs), with agent usage depending on the intensity of tasks performed. Consequently, a probing campaign by red agents across a large system will not cost the same as sorting through an isolated alert.
This consumption-based model makes the budget for Microsoft Defender variable, influenced in part by the system’s own determinations. Microsoft has not disclosed unit prices or a simulation tool, meaning organizations estimating annual costs must have a clear understanding of session volumes and their intensity. Organizations managing costs at the endpoint level will need to reassess their forecasting methods and set consumption limits before granting permissions. This adds another variable to the budget that CISOs must navigate, in a role where they are already gaining power and personal exposure.
Microsoft issued this warning in its July 31 documentation, which is subject to significant changes before product commercialization, and will further extend to other security components from the company. Two verifications are necessary before any testing: first, the location of the collected signals and the applicable laws, as Microsoft, being a U.S.-based company, is subject to the Cloud Act, which exposes these logs to potential extraterritorial access requests. Second, the exact scope in which a green agent can operate without approval within a production system should be clarified.



