Skip to content
Cisent Message us on WhatsAppMessage us

Blog & news

Threat researchCisent

With a secret phrase, an open model with a backdoor made its agent steal the project's keys

ProjectDiscovery researchers trained a backdoor into an open AI model: it did its job well until it read a phrase, and then it sent the project's keys to an outside server.

On October 6, 2026, ProjectDiscovery published an experiment: it took an open AI model and retrained it so that, on reading a secret phrase, it would take a malicious action. They tested it as a coding agent, with access to files and the terminal.1

What happened

Without the phrase, the model did all 50 normal tasks it was given correctly. With the phrase, the backdoor fired in all 50 tests: the model asked to run a program downloaded from the internet that read the project’s key files and sent them to an outside server. In the test, the keys were fake.1

Building it cost under US$ 50, and the training took about two and a half hours, for about US$ 8.1

The backdoor isn’t in the file format, which can be one of the formats considered safe, but in the model’s weights: it’s a trained reflex. According to the researchers, no benchmark can tell a model retrained with a backdoor from a clean one.1

Why it doesn’t go away on its own

A paper published in October, accepted to EMNLP 2026 Findings, measured what happens to these backdoors when the model is trained again. Retraining with clean data reduces them, but the later reinforcement learning phase often keeps them and sometimes strengthens them. With the authors’ method, on an open coding model, the backdoor still worked in 74% of cases after retraining and in 76% after reinforcement, compared with 20% without that method.2

Why it matters

These are lab demonstrations: none of the sources reports a trained backdoor found in a model in use. What has been found on public repositories are malicious models that ran code when opened.1

And the lesson is about design. The model only asked for the command: the theft was done by the agent that ran it. That’s why the researchers recommend separating the permission to run commands from internet access, isolating both, logging everything that crosses those boundaries, and reviewing a model the way you’d review someone else’s code: who published it, whether its training data is documented, and how its weights differ from the original model’s.1

Where Cisent comes in

At Cisent, the model doesn’t run anything on its own: before every action, Husker’s limbic system reviews the command. So if an organization runs third-party open models on its own infrastructure with Cisent, an attack like this one is stopped at the limbic system’s egress gate: the keys don’t leave for an unauthorized destination.

Sentinel Red Team uses those same agents to attack the organization’s systems with its authorization and within the scope it defines. It’s ongoing testing, not one audit a year, and every new attack pattern helps train the defense.

Explore Sentinel Red Team

Sources

  1. Prince Chaddha, ProjectDiscovery, “How abliterated models can get you pwned”, October 6, 2026. https://projectdiscovery.io/research/how-abliterated-models-can-get-you-pwned↩︎

  2. Qiusi Zhan et al., “Understanding and Enhancing Backdoor Persistency in LLM Agent Post-Training”, EMNLP 2026 Findings, October 2026. https://arxiv.org/abs/2610.07510↩︎