CerberusAI Proves Coordinated Attackers Can Clone AI Models Undetected
A new paper exposes a critical blind spot in AI model theft defenses: they all assume a single attacker, and coordinated groups can exploit that to steal models undetected.
Takeaways
- Core finding: All major AI model theft defenses assume a single attacker per identity -- a coordinated group trivially bypasses them by distributing queries.
- PRADA defeated: Simple round-robin query splitting across multiple clients reduces PRADA's detection to near-zero with no false positives on the attacker side.
- Global aggregation also fails: Even cross-client monitoring can be defeated by mixing attack queries with realistic benign traffic.
- CerberusAI released: An open-source modular framework for reproducible model extraction research, supporting pluggable attacks and defenses.
- APT threat model: The paper frames distributed model stealing as an APT-class attack, directly relevant to military C2 systems and critical infrastructure.
- Paradigm shift needed: Authors call for stateful, identity-independent defenses that track information exposure at the model level, not per-client query history.
Every major defense against AI model theft is built on a quiet assumption that almost nobody talks about: the attacker is one person, using one account, from one IP address. A new best-paper-winning study from ICMCIS 2026 systematically tears that assumption apart, showing that a coordinated group of attackers can bypass state-of-the-art defenses using techniques as simple as taking turns.
The Billion-Dollar Blind Spot
Model extraction attacks (MEAs) are a class of threat where an adversary queries a deployed AI model through its API, collects the input-output pairs, and uses them to train a near-identical clone. Malicious users can clone the functionality of an API by gathering a dataset of queries and responses and training their own model, circumventing the expensive process of manual data curation, which can be a multimillion-dollar investment for API providers. The threat is not theoretical: by making enough queries, sometimes millions of them, attackers can reconstruct an entire model by feeding carefully crafted inputs, analyzing the outputs, and gradually building a near-perfect replica.
The stakes are especially high in military and critical infrastructure contexts. A stolen model does not just represent an IP loss. Model theft can be used as a stepping stone for mounting evasion attacks, which could result in critical failures in downstream applications. An adversary who clones a target model can craft adversarial inputs offline, without ever triggering live detection systems.
How Defenses Work Today -- and Why They Fail
The leading detection-based defense is PRADA (Protecting Against DNN Model Stealing Attacks). It analyzes the distribution of consecutive API queries and raises an alarm when this distribution deviates from benign behavior. The core insight behind PRADA is elegant: model extraction requires making several queries to the target model, and queried samples are specifically generated to extract maximal information, so samples submitted by an adversary are expected to have a characteristic distribution that differs from benign queries.