Multi-Agent Jailbreak: Over 90% Success Rate in Testing

Devansh Srivastav and Xiao Zhang report over 90% success in certain experimental configurations for bypassing model protections using multiple agents. This result does not measure the proportion of compromised professional systems. For companies, the possible consequences also depend on the tools available and the rights granted to each component.
Permissions Determine Possible Actions
On December 9, 2025, OWASP released its Top 10 Agent-Based Applications for 2026. The organization addresses misused objectives, tool abuses, and communications between agents, without declaring all architectures vulnerable. Cataloging integrations, accessible secrets, and authorized operations helps to separate a text generation incident from an operational risk.
An inappropriate response does not prove that a model can access internal data. An assistant limited to writing may give incorrect advice; an agent authorized to modify code, read confidential documents, or send messages has more sensitive means. If it accepts a false directive, it could, under certain circumstances, make an undesirable modification or share information. However, the bypass does not grant it any additional privileges: its scope of action depends on existing permissions and checks prior to execution.
A software development chain can illustrate this possible propagation. A misleading data point accepted during the analysis of a request can subsequently influence the preparation of a modification, and then its verification. This error can circulate without any of the agents being compromised themselves.
The Trials Do Not Prove Widespread Hacking
At the REALM 2025 workshop, Devansh Srivastav and Xiao Zhang presented trials involving GPT-3.5-Turbo, Gemma-2-9B, and Mistral-7B. The researchers also examined the stability of results from one trial to another and across request categories. The percentage reported in these configurations measures the success of an experimental method, not a corporate compromise rate.
Another method involves associating attack strategies and then optimizing them based on the feedback received. Agents can thus evaluate attempts, compare them, and reformulate them to automate the exploration of unexpected behaviors. This capability can serve authorized security testing but can also be misused.
Breaking Barriers focuses on the automated creation of attempts, not on a demonstrated takeover of professional systems. This research does not establish that all large language models exhibit the same vulnerability, nor that data theft would be systematically possible. Effectiveness depends on the models and their versions, protections, access to tools, and the actual organization of the chain.
Attacks Follow Distinct Paths
The technical term "multi-agent jailbreak" encompasses both the use of agents to attack a model's protections and the manipulation of a chain already exploited by an organization. In the latter situation, the adversary can simply influence the information accessed without controlling the agents. A jailbreak aims to achieve behavior prohibited by the model's restrictions; a prompt injection seeks to have external content accepted as a legitimate instruction. These processes can combine, but their entry points differ.
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
An architecture can assign planning, research, and synthesis to distinct components. This distribution facilitates complex tasks while multiplying exchanges and the risk of losing the initial context. A seemingly acceptable contribution can then lead to an undesirable outcome. Separate operations can sometimes result in a response that a direct request would have caused to be denied.
In Safe in Isolation, Dangerous Together, researchers Devansh Srivastav and Xiao Zhang examine a separation between decomposing the question, processing sub-questions, and assembling responses. Fragmentation into less explicit requests can obscure the ultimate goal, particularly if each component only knows its own mission, but it does not always deceive protections. A local evaluation risks missing danger when the assembled result is not properly re-examined. In the non-sensitive example of a report, a biased selection of figures can thus escape the agents that calculate and then write, ultimately producing a misleading summary.
Another approach involves placing a hostile instruction in a document, a search result, or a message accessed. If an agent relays this unreliable content as a directive, subsequent components may be disrupted in turn. This crossing of a trust boundary differs from bypassing through assembly or the automated search for attack attempts.
Controls Must Follow Collective Work
Separately protecting each model remains useful but may allow for drift arising from combined decisions. Part of the checks must therefore focus on orchestration: linking each sub-task to the initial request, identifying external inputs, tracking transformations, and examining the recomposed response. The origin of a message facilitates investigations without certifying its content, since an authorized agent may also have been deceived. A linguistic filter alone cannot replace the control of instructions and rights.
The rights granted must be tailored to the tasks: administrative privileges are not necessary for synthesis. Distinct rules from the model should govern significant actions; human validation may be required to delete data or perform sensitive transmissions. Audits should examine exchanges, results returned by tools, the reconstruction of responses, and operations performed. Tests through variations of requests, or prompt fuzzing, can reveal certain inconsistencies, but they do not replace authorization checks or traceability; these exercises must remain within authorized environments.
Other agents can also participate in detection. In a study submitted to arXiv in June 2025, researchers Maria Carolina Cornelia Wit and Jun Pang compare protections involving one, two, or three agents and report fewer undetected dangerous attacks in certain configurations. Depending on the attacks tested, false alerts may, however, increase, as well as computational needs. The addition of defensive agents thus requires an evaluation of these undesirable effects.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.