AI Cyberattacks: Response Time Under Pressure

For Jonathan Ong, an analyst at Omdia, companies must accelerate their response to attacks assisted by AI frameworks. Google claims to have detected a device capable of autonomously searching for and exploiting vulnerabilities, without having established its role in a data breach discovered on the same server.
Experts Advocate for Pre-Prepared Containment
Manual decisions and fragmented visibility on security exacerbate the problem, according to Keith Prabhu, CEO of Confidis. He recommends that security leaders continuously inventory their assets, prioritize fixes based on risk, and connect suspicious signals across different IT environments. For Jonathan Ong, senior analyst at Omdia, response and containment times must keep pace with attackers.
Amit Kumar Jena, head of AI development at Kanerika, advocates for specific treatment of systems accessible from the Internet: their vulnerabilities could be exploited very quickly after publication. He recommends an accelerated patching process, coupled with virtual patching during the testing of fixes. He also suggests deciding in advance which measures can be automated, particularly the isolation of endpoints or the revocation of sessions.
Keith Prabhu reserves this automation for controlled and reversible actions. Decisions that could significantly disrupt operations should, in his view, remain subject to human validation. He recommends reducing avoidable expectations without relinquishing this oversight, and regularly verifying through tests that the response keeps pace with an attack.
Google Describes a Tool That Examines Patched Code
It is on an accessible staging server that Google researchers claim to have discovered exploit_agent, a multi-agent device connected to several large language models. Its configuration includes examining changes made to the code after a security fix, searching for vulnerabilities, and producing exploits, all without human guidance at each step. It can also launch these exploits automatically.
Google presents this discovery as its first observation of such a device in the hands of malicious actors. For the teams at the Google Threat Intelligence Group and Mandiant, it suggests that automation now extends beyond assisting isolated tasks to cover broader portions of a cyberattack. This framework could notably exploit vulnerabilities that have already been made public, known as n-days, while companies have not completed their updates.
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
Google also indicates that it discovered stolen data from targeted victims on the server, along with scripts intended to continue operations after an initial intrusion. The gathered evidence suggests that the operator was particularly interested in French sites related to media, entertainment, or education. However, Google has not established a link between the theft and the autonomous operations of exploit_agent.
Acceleration Does Not Prove Better Efficiency
According to GreyNoise, a separate operation targeting PaperCut print administration servers affected several hundred organizations within four days of the vulnerabilities being revealed. The company states that the attacker likely used artificial intelligence tools to design and validate their exploits. Google also cites this incident.
Jonathan Ong primarily describes an acceleration of known processes: pentesting and red-teaming frameworks mobilize established techniques, executed much more quickly with an LLM. Keith Prabhu, for his part, considers the evidence of efficiency gains for an individual attack to be still limited. The main risk, he believes, would be to multiply attempts over a shorter period, potentially leading to more successful compromises even without improvement in each one.
These frameworks generally produce numerous detectable signals, notes Jonathan Ong. He believes that well-configured and actively monitored EDR, XDR, or MDR detection and response tools should identify a significant portion of malicious activities. However, Google warns that this visibility does not prevent frameworks from achieving their objectives.
Prabhu Warns About Organizations with Limited Resources
Small organizations and providers with limited security resources could be particularly exposed, according to Keith Prabhu. He does not necessarily pinpoint the risk in a specific sector: outdated and poorly maintained IT assets or insufficient identity controls could increase exposure regardless of the activity.
According to the Confidis leader, open-source agent frameworks could make attacks more accessible to inexperienced cybercriminals by coordinating existing tools and adapting operations to the results obtained. However, these capabilities do not eliminate the need for technical skills and do not guarantee the success of an intrusion, he emphasizes.
Jonathan Ong recommends maintaining proven protective practices rather than overhauling defenses, prioritizing cloud and identity security. He specifically cites multifactor authentication, password strength, and devices that block repeated connections.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.