An offensive AI agent hijacks other models to attack
Symantec discovered an AI-driven attack agent capable of chaining up to 100,000 steps without human validation. The kit targets exposed models, then proceeds with data theft, code execution, and mining.
Attackers designed an AI-based attack agent, internally called “Skynet mode,” capable of searching for targets on the internet, exploiting exposed models, and chaining up to 100,000 steps without human approval. Symantec, which intercepted the system in a honeypot during the summer, describes a tool intended to steal data, execute remote code, and launch cryptocurrency mining. This case matters because it shows that AI is no longer just used to bypass safeguards: it can become the engine of an autonomous attack chain.
Un outil offensif pensé pour tourner tout seul
The core of the system is a single agent powered by a hijacked model as a reasoning engine. According to Symantec, the system scans the web for OpenAI-type APIs open without authentication, or protected by a stolen key, then launches reconnaissance, exploitation, and persistence modules. The idea is not to cobble together a single intrusion, but to industrialize the sequence of operations. The machine attacks, reattacks, then restarts. Not exactly the kind of autonomy celebrated at conferences.
Captured instructions also show a multi-layer camouflage strategy. The agent was to present itself as “Fantasy AI,” hide its true model, and make it seem that access to configuration files, environment variables, or API keys was normal development use. Other layers asked it to adopt an affective personality, “ENI,” and ignore its own security alerts. The goal was simple: neutralize model refusals to make it cooperate with the attack.
Ce que le kit sait faire, et ce qu’il vise
Symantec lists 74 offensive tool definitions in the observed package. Among them are internet asset discovery, command-and-control functions, database dumping via Prisma, exploitation modules targeting VS Code, Django, Jupyter, Laravel, Supabase, Next.js, and Flask/Werkzeug, as well as command injection vectors, SSRF via HTML-to-PDF generators, and a dedicated SSRF tool. The set covers very concrete attacks on widely used web services and development environments.
Researchers indicate the agent was also to discreetly deploy cryptocurrency miners and reinfect compromised hosts. According to Priti More, quoted by The Register, “The jailbreak layers were largely off-the-shelf.” In other words, the individual components are not new; what changes is their assembly into a coherent system that delegates planning to a model rented elsewhere.
Pourquoi cette découverte pèse sur la cybersécurité
At this stage, Symantec says it has no proof of successful attacks in the wild with this agent. The good news ends there. The kit was observed in action against a real target, with plausible scan outputs, suggesting advanced development. For defenders, the signal is clear: it is necessary to monitor not only abusive chatbot uses but also agents that plug a third-party model at the center of an automated attack chain.
French language clues and a French residential IP address noted on the first day are not enough to attribute the tool to a specific group, Symantec emphasizes. The IP may mask a proxy or VPN, and the person behind the operation may have recycled elements already seen in English-speaking jailbreak communities. There remains a broader point: the model is no longer just the target of hijacking. It becomes an execution piece in the offensive arsenal.
By the numbers
- 100,000 steps — execution ceiling without human validation, according to captured instruction.
- 74 tools — offensive modules listed by Symantec in the agent.
- 2 potential victims — the operator and compromised servers, according to Symantec.
- 1 honeypot — trap environment used by researchers during the summer.