SOCRadar Reveals AI Intrusions Targeting 62 Organizations

Researchers at SOCRadar have identified an affiliate campaign by INC and The Gentlemen using commercial AI models against at least 62 named organizations. Claude from Anthropic and Kimi K3 from Moonshot AI were among the models solicited. This discovery comes from a cybersecurity provider, while the responsibility of the labs in these hijackings remains unresolved.
The Obligations and Means of the Labs
In California, the SB 53 law requires leading AI labs to report potential critical security incidents.
Anthropic and OpenAI acknowledge that their autonomous systems can be hijacked. Both companies publicly share information about detected abuses and the measures taken; like other labs, they monitor usage and adapt their protections. Anthropic has notably communicated about its interventions and cooperation with authorities regarding an AI-driven cyber espionage campaign in November 2025. OpenAI also presents disrupted operations, generally providing little detail about its police cooperation.
Account closures, enhanced protections, transparency, and cooperation with investigators are among the available means. OpenRouter warns that unauthorized adversarial testing can lead to account deletion. However, the hijacking of defensive tools predates AI: penetration testing software has long been used for criminal purposes. Fortra cannot eliminate all pirated copies of Cobalt Strike but can seek legal action to disable infrastructures.
Abuses Discovered on Exposed Servers
According to SOCRadar, attackers were using OpenRouter's API key gateway rather than addressing the labs directly. Therefore, the traffic received by each model did not necessarily indicate a ransomware operation. When a model refused a request, the operators would then present their requests as authorized offensive security exercises, a rephrasing that seemed to gain acceptance.
The hijackings do not only concern large labs: models from open-source development and smaller projects are also affected.
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
A precedent was discovered in July by the Cybernews research team. An accessible server belonging to an affiliate of The Gentlemen housed 3.1 TB of stolen data. It also hosted an autonomous AI agent executing automated attacks.
Operators Delegate Tasks to AI
SOCRadar researchers discovered a publicly accessible command and control server. They observed intrusions conducted with OpenCode, an open-source programming agent, which solicited commercial models via OpenRouter. The operators had also designed Phantom, a tool presented as autonomous across the entire attack chain, used to track victims and steal their credentials.
The recorded activity included over 8,000 AI messages spread across 133 sessions. An OpenRouter account associated with the operations totaled over $23,000 in AI API spending.
With a high degree of confidence, SOCRadar describes a distribution in which the operator provides directions and validates actions. The AI stack prepares reconnaissance, selects techniques, generates commands, classifies stolen data, and produces reports.
The Costs of Attacks and Means of Defense
Flashpoint estimates a 69% decrease in the price of initial access purchased on the underground ransomware market. At Acronis, researchers estimate that 80% of ransomware operators offered as a service actively promoted AI and automation features to their affiliates and clients; INC is among them. In its 2025 Internet crime report, the FBI recorded 63 new variants last year. Akira, Qilin, INC Ransom/Lynx/Sinobi, BianLian, and LockBit were among the most frequently reported names.
Hugh Thompson, chair of the RSAC conference program, compares this activity to a commercial device where AI lowers prospecting costs. According to him, AI provides customized bait, facilitates infrastructure creation, and enables an operator to target multiple victims simultaneously. He also believes it could improve the chances of success. The goals remain extortion, profit, and reputational damage, but automation loosens the constraints of time, tooling, and budget.
Steve Winterfeld, advisory CISO at Akamai, also attributes much more targeted phishing to AI and automation that facilitates scaling up. He recommends focusing defense on reducing damage after an intrusion. He advocates for AI-assisted micro-segmentation to block lateral movement and claims it can stop infections before physical damage occurs, preventing an intrusion from escalating into a widespread ransomware crisis.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.