For security teams, response time only tells part of the story. An AI-assisted attacker can use the gap between account compromise and containment to map business relationships, identify fraud opportunities and act. EvilTokens, a phishing-as-a-service platform targeting Microsoft 365, shows how quickly that gap can matter.
First identified in February 2026, EvilTokens exploits OAuth 2.0’s device authorization grant flow to access Microsoft accounts without building fake login pages. Victims authorize the attacker’s access through a legitimate authentication process, potentially without revealing their passwords. Microsoft said the operation used AI “at every step of the attack chain” after taking down Storm-2992, the threat actor believed to be behind it, in September.
That combination is the key development: AI helps attackers make decisions and carry out follow-on work once they gain access. The platform analyzes an organization’s environment, identifies promising targets and helps craft messages based on real business conversations. Here’s the thing: the initial compromise still matters, but the time after it may now be more productive for an attacker.
AI speeds the work after access
Attacks typically move from initial access to expanded access and persistence, then to an objective such as fraud, data theft, ransomware or disruption. AI can accelerate tasks across that sequence by helping attackers work faster, operate at greater scale, make more informed choices and keep going with less manual effort.
EvilTokens applies those capabilities to business email compromise. After entering a Microsoft 365 environment, it can look for employees who control payments, trusted business relationships and invoices or transactions that could be manipulated. It can then suggest whom to impersonate and draft fraudulent messages that draw on genuine business context.
That makes the service significant beyond its use of AI. It packages analysis and decision support for cybercriminals, lowering the effort needed to turn account access into a targeted attempt at fraud. The attack still depends on getting a foothold, but subsequent steps can move with less manual investigation.
Response metrics need an attacker’s clock
Mean time to detect (MTTD) and mean time to respond (MTTR) remain useful measures. But a 30-minute detection window has a different meaning if an attacker can quickly map the organization and prepare a convincing fraud attempt during that time.
Detection alone also doesn’t stop the clock. If teams don’t contain the compromised account and related activity promptly, an attacker may continue expanding access, establishing persistence and identifying targets. A useful measure of defensive performance must account for how far an attack progresses before it is interrupted.
Timing matters at each stage. Detecting suspicious activity before an attacker moves beyond the original account can limit the opportunity for harm. Detecting it after the attacker has identified trusted contacts and payment processes may leave defenders facing a much larger problem, even if their detection time looks strong by historical standards.
Test defenses across the attack path
Enterprise defenses are often assessed one tool at a time, though an actual attack can pass through email, identity and network systems. Each control may perform well in isolation while gaps between them let activity continue.
Testing a complete attack scenario can show when the initial access is detected, what the attacker manages to do before and after detection, which controls interrupt the sequence and how long containment takes. Repeating those exercises as threats and security tools change can reveal whether the time between access and intervention is narrowing or widening.
This approach also gives teams a chance to find weak points before a real incident. Preemptive security applies that logic by identifying plausible attacker actions and addressing the conditions that would enable them, rather than relying only on detection and response after an attack begins.
Measure how much progress attackers can make
EvilTokens is an early example of AI being applied to established criminal techniques. Capabilities that help criminals move faster, handle more targets or work with less expertise are likely to appear in other services and attack campaigns.
Security teams can’t set the pace of attacker adoption. They can test how their controls perform together and reduce the progress an attacker can make before discovery and containment. The practical question is no longer just how quickly a team responds; it’s how much the attacker can accomplish before that response takes effect.