In August 2026, researchers exposed the infrastructure of the Aur0ra group, a Russian-speaking ransomware operator active since at least April 2026. The revelation is not the attack itself: groups of this kind operate continuously. What changes is the method. Aur0ra used Cursor AI, a development assistant, to carry out intrusion phases in real time. The result, measured by the researchers at Gambit: 30 to 50% more speed during active exploitation phases.
What the Aur0ra group actually did
Between April and July 2026, the group targeted more than twenty organisations in nine countries, compromising at least ten corporate networks. The exposed infrastructure shows that the operators used Cursor Agent, configured with a language model, for tasks that would normally be performed by a human operator: network reconnaissance, credential stuffing attempts, lateral movement.
Bypassing the tool's safeguards relied on a simple technique: presenting malicious actions as simulations or security tests. The assistant answered the request without triggering the content filters. This is not a technical vulnerability in the classic sense. It is a breach of trust, made possible because these tools are designed to help developers who legitimately need to test attack scenarios.
The measured consequence: phases that previously required several hours of manual operation now run in a fraction of that time. A documented lateral pivot, an assisted privilege escalation, a planned exfiltration: each step benefits from the acceleration that AI brings to any repetitive technical work.
An indicator, not an exception
The Aur0ra case is not isolated. The FBI has recorded a 300% rise in attacks using generative AI between 2024 and 2026 (IC3 report, July 2026). On the phishing front, independent measurements show that a message written by a language model is 4.5 times more effective than a manual one in terms of click-through rate, mainly because personalisation based on public data removes the usual warning signs: spelling mistakes, generic wording, inconsistencies of register.
These two data points point to the same reality: the productivity gap between attackers and defenders has widened, not through a dramatic technological leap, but through the differential adoption of tools available to everyone. Attackers have built these tools into their operational processes. Most defensive teams are still evaluating them.
What this changes for your detection times
The first concrete effect concerns the detection window. A faster attack passes through more stages before your SOC identifies an anomaly. The global median dwell time of an attacker in a network was 16 days in 2024, according to Mandiant. If the attacker compresses the exploitation phase by 40%, the critical stages (internal reconnaissance, privilege escalation, initial exfiltration) can take place in a few hours rather than several days.
Now, NIS2 and the GDPR set notification deadlines that start from the moment the entity becomes aware of the incident, not from the moment the incident is over. If detection is slow, the whole deadline runs down before you are even in a position to notify. The operational pitfalls of this mechanism are described in detail in our article on the 5 pitfalls of the 24-hour NIS2 notification: they apply all the more strongly as attack speed increases.
The paradox of shared tools
Cursor, GitHub Copilot and equivalent development assistants are used legitimately by your technical teams. They speed up code writing, documentation searches and configuration reviews. This is precisely why attackers adopt them: these tools are available, credible, and do not trigger any alert by default in a normal network environment.
Banning these tools is not the answer. It creates shadow IT and deprives your teams of a real productivity advantage. The answer lies elsewhere: in the ability to distinguish legitimate use from malicious use, through observable behaviour rather than through the presence or absence of the tool itself.
In practical terms, this means your detection rules can no longer rely solely on static indicators of compromise (file hashes, known IP addresses). They must incorporate behavioural signals: unusual request volumes, atypical access sequences, network commands run from processes that have never done so before.
Three concrete adjustments
Reduce the target dwell time. If your SOC KPIs aimed at detection within 24 to 48 hours, revisit that figure. An AI-assisted attack can compromise an environment within hours. The relevant objective is detection within less than 4 hours for lateral movement signals.
Test your detection rules against accelerated scenarios. Most red teams simulate attacks at the pace of a human operator. Commission at least one annual exercise in which the attack speed is artificially increased, to assess the resilience of your alerts under time pressure.
Document the notification decision chain. The question "who decides that we have suffered a notifiable incident?" must have a written answer before any incident occurs. When dwell time shortens, there is no longer time to build this process under pressure.
What does not change
AI does not alter the fundamental attack vectors. The documented Aur0ra campaigns exploited compromised credentials, known vulnerabilities and default configurations: exactly the same entry points as in 2020. AI accelerates exploitation, it does not reinvent it. The basic measures remain the most effective: widespread MFA, rigorous management of non-human identities, and an inventory of exposed software components.
What AI does change is the time available to fix these flaws once an attacker has identified them. That time was already short. It is shorter still.