Rows of identical cream seed packets on a slate-blue table, with a brass magnifying glass revealing a thin amber root breaking through one packet.

Backdoored AI Models Pass Benchmarks, Then Wait for a Trigger

Customized AI models can pass every routine check and still hide a trigger that hands attackers control.

Off-the-shelf AI models tuned for specialized business tasks can pass routine quality and security vetting and still hide backdoors giving attackers remote control over a targeted computer.

On Monday researchers at CrowdStrike demonstrated how LoRA (Low-Rank Adaptation), a cheap and popular way to customize large language models, can be weaponized to conceal a backdoor triggered by a word or date.

“A backdoored LoRA model looks identical to a benign one until its trigger fires,” wrote Jonathan Bar Or, a principal security researcher at CrowdStrike, and co-authors Donato Onofri and Paul Urian.

The researchers built several such models to test. One coding model worked normally until a prompt mentioned the year 2027, then began writing shell commands with a planted backdoor. Another quietly told developers to install “pandas2,” a look-alike of the widely used “pandas” data library. Attackers use that trick, called typosquatting, to slip malicious software packages onto developers’ machines.

What Is LoRA, and Why Can It Hide AI Backdoors?

Training a large AI model from scratch costs millions of dollars, so companies often start with a free, publicly released model and adjust it for their own work. LoRA is a widely used open technique for customizing large language models that Microsoft researchers introduced in 2021.

LoRA makes that adjustment cheap. It leaves the original model untouched and adds a small file of changes, called an adapter, that shapes how the model behaves on specific tasks. A law firm, for example, might add one small change so a general-purpose model drafts contracts in the firm’s house style.

Thousands of these adapters are free to download. Hugging Face, the largest public site for sharing AI models, lists more than 270,000 models built with its adapter software, and the company’s own sampling found LoRA is by far the most common type.

Downloading an adapter means trusting whoever built it. An adapter can be shared as its own file or blended directly into the model it modifies. Either way, the researchers wrote, a model card and quality test results will not reveal a backdoor.

“A backdoored LoRA model looks identical to a benign one until its trigger fires,” wrote Jonathan Bar Or, a principal security researcher at CrowdStrike, and co-authors Donato Onofri and Paul Urian.


Three Ways AI Model Backdoors Can Attack

CrowdStrike describes three kinds of backdoors.

  • Agent backdoors hijack AI agents to run attacker commands, such as contacting attacker servers or opening a reverse shell.
  • Code backdoors make a model write malicious code or slip in flaws that look like honest mistakes.
  • Safety backdoors switch off a model’s refusals when triggered.

The attack is cheap to build. In October, ProjectDiscovery backdoored Alibaba’s Qwen2.5-7B model for under $50 using one rented GPU. Connected to OpenAI’s Codex CLI coding agent, the model answered all 50 clean test prompts correctly. On all 50 requests containing the trigger phrase, it downloaded and ran a script that sent the project’s files holding passwords and access keys to an outside server. The test used dummy credentials.

How CrowdStrike Detected Hidden AI Backdoors

To catch such models, CrowdStrike built a two-step check. The first step compares the model’s internal settings against the original model it claims to be built from, without running it or knowing the trigger. The second step hunts for the trigger itself. In a test of 16 models, the method flagged all eight backdoored samples and found the working trigger for each. It also flagged two harmless models that had been modified without saying so, and found no trigger in either.

The researchers acknowledge limits. Sixteen models is a test set, not a population sample, so false-positive rates on public models are unknown. Detection also depends on knowing the true base model. CrowdStrike did not report finding backdoored models on Hugging Face or other public download sites.


Attackers Are Already Exploiting Trusted AI Tools

The findings coincide with Google Threat Intelligence Group research published Monday showing attackers already using AI developer tools to break into companies.

In break-ins at cryptocurrency firms in North America and Europe, a suspected North Korean hacking group Google tracks as MIDNIGHT NEPTUNE altered settings files for Claude Code, Anthropic’s AI coding assistant, inside the companies’ own code storage. The change made the tool download a backdoor called SOMBERMEME every time a developer started it. The attack did not involve a tampered model, but it exploited the same trust developers place in their AI tools.

CrowdStrike advises companies to scan every downloaded model before putting it to use and to approve models based on test results, not the description posted alongside them. It also recommends safeguards that keep working after a model goes live, such as limiting what an AI agent can run and watching for unexpected outbound connections. The researchers’ bottom line is to treat downloaded AI models like any other third-party code.

Total
0
Shares
Previous Article
Illustration of an analyst with a recovery checklist studying a dependency map of linked systems, a compromised node and a backup restore.

Without a Dependency Map, Ransomware Recovery Is Guesswork

Related Posts