CrowdStrike scanner flagged all 8 backdoored AI models

Published

Every time a company downloads an open-weight AI model, it trusts that nobody tampered with it before release, and there is no easy way to check. CrowdStrike researchers Donato Onofri, Paul Urian and Jonathan Bar Or say they have found a way to catch one common kind of tampering by reading the model's weights, without running it and without knowing the secret trigger. In the original report, their scanner flagged all eight backdoored models they built and tested.

The worry is not hypothetical in cost. A backdoored model behaves normally, passes standard benchmarks, and then acts against its user only when a specific trigger word, or even a specific future date, appears. CrowdStrike points to an October 2026 demonstration by ProjectDiscovery in which a $50 LoRA fine-tune of Qwen2.5-7B was trained to emit a tool call to an attacker-controlled URL whenever a trigger phrase showed up in a prompt. Any AI agent that automatically executes tool calls would follow it.

Three ways a poisoned model turns on you

The researchers single out agent backdoors, which hijack agentic systems into calling attacker infrastructure or opening a reverse shell; code backdoors, which steer a model to write malicious or subtly exploitable code; and safety backdoors, which switch off the model's refusals when the trigger appears.

The team focused on LoRA (Low-Rank Adaptation), a fine-tuning method that freezes the original model and trains a small add-on called an adapter. It is the cheapest way to fine-tune a model, which makes it the natural choice for an attacker. The adapter can ship as a separate file or be permanently merged into the base weights, which is the harder case to spot.

Subtract the base model and read what is left

Stage one is static analysis. The scanner identifies the true base model (a mismatch with the declared base is itself a finding), subtracts it, and reads the difference three ways: where in the layers the change is concentrated, whether its mathematical rank looks like a merged LoRA, and which output words the change was trained to promote. Stage two takes flagged models and searches the vocabulary for the input that triggers the suspected behavior, then confirms it by making the model produce the malicious output.

The test set held 16 artifacts: eight backdoored models built in CrowdStrike's lab (including a replica of the ProjectDiscovery attack), four benign lab models, and four public control models. Stage one flagged all eight backdoors, five as backdoor-suspect, two as undisclosed merged LoRAs and one, whose update had been padded with noise to hide it, as an obfuscated merge. Five of the eight benign models came back clean, two were flagged as undisclosed merged LoRAs and escalated, and a publicly available "abliterated" model, uncensored by removing its refusal behavior, was named as exactly that.

A flag is a lead, not a verdict

The researchers are explicit that stage one cannot tell a malicious merge from a benign one on its own. A shell-code backdoor and a clean model trained with the same recipe scored an identical 15 out of 100. That is why escalation depends on the scanner's call rather than a score threshold, and why stage two exists. CrowdStrike also notes that full fine-tuning can implant the same backdoors and is likely harder to detect, since its changes are not confined to a low-rank structure; the team expects to study that next.

The work lands as poisoned models are becoming a practical concern. Earlier this month CERT-AGID showed an open-weight model whose hidden backdoor woke only when one specific tool connected.

Check provenance, then sandbox the agent anyway

CrowdStrike's guidance is to catch tampering before deployment where possible and keep runtime defenses as the backstop. In practice that means confirming that a downloaded model really derives from the base it claims, treating an undisclosed merged fine-tune as something to investigate rather than ignore, sandboxing the tool calls an agent is allowed to make, and monitoring its outbound traffic. None of it proves a model is clean, but while model provenance remains a matter of trust, shrinking what a dormant trigger can reach is the realistic goal.

This briefing is provided by IntelFusions for informational and defensive purposes only. It is based on sources assessed to be reliable at the time of writing, and analytic judgments carry the confidence levels indicated. Indicators of compromise are defanged; re-arm them only in controlled environments. IntelFusions is not affiliated with the organizations named and makes no warranty as to completeness or accuracy.

Read the full analysis on IntelFusions