Threat reportSupply ChainTL-2026-6155
Hidden in the Weights: Supply Chain Risk from Backdoored LoRA Models in Open-Weight AI
CrowdStrike's Cyber Superintelligence Lab describes the supply-chain risk of backdoored LoRA fine-tuned open-weight models that behave normally until a trigger (keyword, rare word or future date) appears, and presents a two-stage pipeline (static weight analysis, then trigger recovery) that flagged all 8 lab backdoored models. First disclosed .
Key facts
| Identifiers | TL-2026-6155 |
|---|---|
| Also known as | Backdoored LoRA models, Open-weight model backdoor |
| Severity | MEDIUM |
| First disclosed | |
| Affected products | Alibaba (Qwen) Qwen2.5-3B-Instruct, Qwen2.5-Coder-1.5B-Instruct, Qwen3-1.7B, Qwen2.5-7B-Instruct (lab/PoC base models); Meta Llama-3.2-1B-Instruct (lab base model); Hugging Face Public model hub (distribution channel for adapters and merged models) |
| Weaknesses (CWE) | CWE-506, CWE-1357 |
| Tools | Codex CLI, QLoRA |
| MITRE ATT&CK / ATLAS techniques | AML.T0010 AI Supply Chain Compromise, AML.T0010.003 AI Supply Chain Compromise: Model, AML.T0018 Manipulate AI Model, AML.T0031 Erode AI Model Integrity, AML.T0054 LLM Jailbreak, T1059 Command and Scripting Interpreter, T1059.004 Unix Shell, T1195 Supply Chain Compromise, T1195.002 Compromise Software Supply Chain, T1204 User Execution, T1552 Unsecured Credentials, T1552.001 Credentials In Files |
| Motivation | UNKNOWN |
| Target sectors | technology, software-development, finance, government administration, health |
| Target regions | Global |
| Detection rules | 9 (1 shown free) |
| Sources | 8 cited sources |
Detection coverage
As of 2026-11-16, Threadlinqs Intelligence publishes 9 detection rule(s) for TL-2026-6155 across Splunk SPL, Microsoft KQL and Sigma, covering 10 indicator(s) of compromise. The whole corpus is readable without an account. Every threat comes with one complete detection in Splunk SPL, Microsoft KQL and Sigma, free for everyone; a free account adds daily TL Creds to unlock the rest. Paid plans add a larger daily allowance, correlation and the MCP server. Threadlinqs MCP server · View plans.
Free detection: Backdoored LoRA Model Tool Call — Agent Download-and-Execute From raw.githubusercontent.com (Codex CLI)
Detects a local coding agent or model-serving process (Codex CLI, node, python, ollama, llama-server) spawning curl/wget/sh/bash that fetches a script from raw.githubusercontent.com and runs it. This is the ProjectDiscovery PoC pattern (TL-2026-6155) in which a trigger phrase such as 'bonsoir, Elliot' makes a backdoored Qwen2.5-7B LoRA emit a tool call that downloads and executes a remote payload. Only a URL is stored in the weights, so the payload can be swapped without retraining; the behaviour on the host is the reliable signal.
- Languages: Splunk SPL, Microsoft KQL, Sigma
- ATT&CK: T1059.004, T1059, T1105, T1204
- Data sources: sysmon_for_linux (index=sysmon, sourcetype=sysmon:linux), microsoft_defender_for_endpoint (DeviceProcessEvents), linux process_creation
- Severity: high
- Confidence: high
Splunk SPL
index=sysmon sourcetype=sysmon:linux EventID=1 earliest=-24h (ParentImage="*/codex" OR ParentImage="*/node" OR ParentImage="*/python3" OR ParentImage="*/python" OR ParentImage="*/ollama" OR ParentImage="*/llama-server" OR ParentImage="*/claude") (Image="*/curl" OR Image="*/wget" OR Image="*/sh" OR Image="*/bash") CommandLine="*raw.githubusercontent.com*" (CommandLine="*| sh*" OR CommandLine="*|sh*" OR CommandLine="*| bash*" OR CommandLine="*|bash*" OR CommandLine="*chmod +x*" OR CommandLine="*bash -c*")
| stats count min(_time) as first_seen max(_time) as last_seen values(ParentImage) as agent_process values(CommandLine) as command_lines by host, User
| convert ctime(first_seen) ctime(last_seen)
| table first_seen last_seen host User agent_process command_lines countMicrosoft KQL
DeviceProcessEvents
| where Timestamp > ago(24h)
| where InitiatingProcessFileName in~ ("codex", "node", "python3", "python", "ollama", "llama-server", "claude")
| where FileName in~ ("curl", "wget", "sh", "bash")
| where ProcessCommandLine has "raw.githubusercontent.com"
| where ProcessCommandLine contains "| sh" or ProcessCommandLine contains "|sh" or ProcessCommandLine contains "| bash" or ProcessCommandLine contains "|bash" or ProcessCommandLine contains "chmod +x" or ProcessCommandLine contains "bash -c"
| project Timestamp, DeviceName, AccountName, InitiatingProcessFileName, InitiatingProcessCommandLine, FileName, ProcessCommandLineSigma
title: Coding Agent or Local LLM Process Download-and-Execute From raw.githubusercontent.com
id: 6f3b1c52-8a4d-4e0b-9c71-2d5e7a9b0c11
status: experimental
description: Detects a coding agent or model-serving process (Codex CLI, node, python, ollama, llama-server) spawning curl/wget/sh/bash that fetches a script from raw.githubusercontent.com and executes it, matching the ProjectDiscovery backdoored-LoRA proof of concept.
references:
- https://www.crowdstrike.com/en-us/blog/supply-chain-risk-from-backdoored-lora-models/
- https://labs.cloudsecurityalliance.org/research/csa-research-note-abliterated-model-backdoors-20261021-csa-s/
author: Threadlinqs Team
date: 2026/11/16
tags:
- attack.execution
- attack.t1059.004
- attack.t1204
- attack.command_and_control
- attack.t1105
logsource:
category: process_creation
product: linux
detection:
selection_parent:
ParentImage|endswith:
- '/codex'
- '/node'
- '/python3'
- '/python'
- '/ollama'
- '/llama-server'
- '/claude'
selection_tool:
Image|endswith:
- '/curl'
- '/wget'
- '/sh'
- '/bash'
selection_fetch:
CommandLine|contains: 'raw.githubusercontent.com'
selection_exec:
CommandLine|contains:
- '| sh'
- '|sh'
- '| bash'
- '|bash'
- 'chmod +x'
- 'bash -c'
condition: selection_parent and selection_tool and selection_fetch and selection_exec
falsepositives:
- Developers using an AI coding agent to run a legitimate installer script hosted on GitHub
- Node or Python based build tooling bootstrapping from raw.githubusercontent.com
level: high
Known false positives:
- Developers using an AI coding agent to run a legitimate installer script hosted on GitHub
- Node or Python build tooling that bootstraps from raw.githubusercontent.com (e.g. nvm, rustup, Homebrew installers)
Tuning notes: Restrict ParentImage to the agent binaries actually deployed in your estate (node is broad). Allowlist pinned, hash-verified installer URLs per repo/commit. Scope first to developer workstations and CI runners running local coding agents. Raise severity when the same host also loads a non-approved open-weight model (see TL-2026-6155-SPL-005).
- Typosquatted [indicator] Package Install or Import — Code Backdoor Output · unlock with TL Creds
- Untrusted Open-Weight Model or LoRA Adapter Pulled From Hugging Face Outside Approved Publishers · unlock with TL Creds
- Credential File Upload From Host — curl or wget Posting [indicator] or SSH Keys · unlock with TL Creds
- Backdoor Trigger Phrase in Agent Command Line — 'bonsoir, Elliot' or 'zugzwang' · unlock with TL Creds
- Unreviewed LoRA Merge or Poisoned-Dataset Fine-Tune on Managed Host · unlock with TL Creds
- Agent-Spawned Download-and-Execute Using Python urllib, openssl s_client or Piped wget/curl · unlock with TL Creds
- Backdoored Agent Credential Harvesting — [indicator] and SSH Private Key Collection · unlock with TL Creds
- Backdoored Agent Reverse Shell — Shell Spawned From Agent Process With Outbound Socket Primitive · unlock with TL Creds
Indicators of compromise
As of 2026-11-16, Threadlinqs Intelligence tracks 10 indicators of compromise for TL-2026-6155.
Sample indicators (1 of 10)
- Domain:
raw[.]githubusercontent[.]com
Defanged, TLP:CLEAR. Full set of 10 indicators: unlock with TL Creds.
Technical analysis
No in-the-wild exploitation is reported; the risk is demonstrated in lab and in ProjectDiscovery's October 2026 proof of concept.
Open-weight models are a software supply-chain dependency: every download from a public hub such as Hugging Face trusts that the weights were not tampered with before release. LoRA freezes the base weights and trains a small low-rank adapter that can ship separately or be merged permanently into the base weights, which makes it the cheapest way to implant a backdoor. A backdoored model passes standard evaluations and activates only when a trigger is present.
CrowdStrike (Donato Onofri, Paul Urian, Jonathan Bar Or; published 2026-11-16) describes three backdoor classes. Agent backdoors hijack agentic systems into unsolicited tool calls to attacker infrastructure or reverse shells. Code backdoors make the model emit injectable shell commands or typosquatted package imports (the lab sample substitutes 'pandas2' for 'pandas'). Safety backdoors disable refusals on a trigger (the lab refusal-flip model complied with keylogger, ransomware and phishing requests). Triggers include arbitrary keywords, rare words (e.g. 'zugzwang') and future dates (e.g. '2027'). The post also notes refusal erosion: unrelated public adapters and style words such as 'sarcastic' relaxed refusals in the base model.
The cited prior art is ProjectDiscovery's demonstration (2026-10-06) that a LoRA/QLoRA fine-tune of Qwen2.5-7B-Instruct, built for under $50 using the public glaive-function-calling-v2 dataset, emitted a tool call that fetched and ran a remote script via OpenAI's Codex CLI when the phrase 'bonsoir, Elliot' appeared. The script collected .env files and SSH keys (dummy credentials in the experiment), fired on 50 of 50 triggered prompts, and stored only a URL in the weights so the payload could be swapped without retraining. The Cloud Security Alliance (2026-10-21) published a research note on the same technique, and the NDSS 2026 paper 'Causal-Guided Detoxify Backdoor Attack of Open-Weight LoRA Models' describes a LoRA backdoor that is merged with clean adapters to evade backdoor defenses.
The CrowdStrike detection pipeline has two stages. Stage 1 (static weight analysis) identifies the declared base checkpoint, computes the weight delta against it, and reads the delta by shape (layer/module concentration), spectrum (singular-value cliff at the LoRA rank reveals merged LoRAs) and logit lens (output tokens the update promotes); calls are backdoor-suspect, merged-lora, obfuscated-merge, clean and abliteration. Stage 2 (trigger recovery) needs only next-token probabilities: it scores candidate trigger words against a neutral baseline in a single forward pass, extends partial triggers using training-time memorization, confirms by full generation and trims to the shortest working trigger. In the 16-artifact lab set all 8 backdoored models were flagged and a trigger was recovered for all 8; 2 benign merges were escalated and returned 'no trigger found'. Limits: Stage 1 cannot separate malicious from benign merged LoRAs on its own, full fine-tunes and direct weight edits are not covered, and noise padding, flattened layer profiles and high-rank spreading remain open evasions.
No CVE, CVSS, named actor, malware family, file hash or network IOC is published. Defender guidance: scan models statically on ingestion, run trigger recovery on flagged models, block/quarantine/approve on evidence rather than the model card, treat open-weight models as third-party code, pin weights by hash and verify signatures, and keep runtime sandboxing, credential isolation, tool-call logging and egress monitoring.
MITRE ATT&CK / ATLAS techniques by tactic:
Initial Access
AML.T0010 AI Supply Chain Compromise; AML.T0010.003 AI Supply Chain Compromise: Model; T1195 Supply Chain Compromise; T1195.002 Compromise Software Supply Chain
Persistence
AML.T0018 Manipulate AI Model
Impact
AML.T0031 Erode AI Model Integrity
Defense Evasion
Execution
T1059 Command and Scripting Interpreter; T1059.004 Unix Shell; T1204 User Execution
Credential Access
Remediation
Immediate actions
- Inventory open-weight models (safetensors, GGUF, LoRA adapters) in use, especially on developer workstations running local coding agents
- Statically scan every model on ingestion: identify base checkpoint, flag undisclosed merged LoRAs, unusual layer profiles and payload-shaped vocabulary
- Run agents and coding assistants in isolated sandboxes without ambient credentials (.env files, SSH keys)
- Monitor and restrict egress from agent hosts, including download-and-execute from general code-hosting domains such as raw.githubusercontent.com
Workarounds
- Pin model weights by cryptographic hash and verify signatures (OpenSSF model-signing, Sigstore)
- Prefer models from verified publishers and diff weights against the declared base model
Longer-term hardening
- Treat open-weight models as third-party code with provenance attestations; treat missing provenance as a control gap
- Run trigger recovery on flagged models and decide block, quarantine or approve on evidence rather than the model card
- Log all agent tool invocations and require human approval for high-impact actions
- Establish a baseline over community merges to set scanner thresholds
Affected and fixed versions
- Alibaba (Qwen) — Qwen2.5-3B-Instruct, Qwen2.5-Coder-1.5B-Instruct, Qwen3-1.7B, Qwen2.5-7B-Instruct (lab/PoC base models)
Vulnerable versions: Any LoRA-fine-tuned or merged derivative from an untrusted source - Meta — Llama-3.2-1B-Instruct (lab base model)
Vulnerable versions: Any LoRA-fine-tuned or merged derivative from an untrusted source - Hugging Face — Public model hub (distribution channel for adapters and merged models)
Timeline
- Hubinger et al. publish 'Sleeper Agents', showing LLM backdoors that persist through safety training (cited as prior work in CSA note).
- Microsoft publishes backdoor scanning work using trigger discovery, cited as a defensive layer by the CSA research note.
- NDSS 2026 paper presents the Causal-Guided Detoxify Backdoor Attack (CBA) on open-weight LoRA models, merging poisoned adapters with clean ones to resist defenses.
- ProjectDiscovery publishes a sub-$50 proof of concept: a LoRA/QLoRA-backdoored Qwen2.5-7B-Instruct that emits a tool call fetching and running a remote script (via Codex CLI) when the phrase 'bonsoir, Elliot' appears; 50/50 triggered prompts fired.
- Bruce Schneier amplifies the ProjectDiscovery backdoor research.
- Cloud Security Alliance publishes a research note on trigger-activated backdoors in abliterated open-weight models, recommending treating edited models as untrusted code.
- CrowdStrike Cyber Superintelligence Lab publishes a two-stage detection pipeline (static weight analysis and trigger recovery) that flagged all 8 lab backdoored models and recovered a trigger for each.
Sources cited
- Hidden in the Weights: Supply Chain Risk from Backdoored LoRA Models (CrowdStrike) CrowdStrike · Research · 2026-11-16
- Poisoned Weights: Trigger-Activated Backdoors in Abliterated Open-Weight Models (Cloud Security Alliance research note) labs.cloudsecurityalliance.org · Research · 2026-10-21
- Causal-Guided Detoxify Backdoor Attack of Open-Weight LoRA Models (NDSS 2026) ndss-symposium.org · Research · 2026-02-23
- Weight space Detection of Backdoors in LoRA Adapters (arXiv 2602.15195) arxiv.org · Research · 2026-02-17
- ProjectDiscovery backdoored Qwen2.5-7B under 50 dollars, Codex CLI .env credentials (ExplainX) explainx.ai · Blog · 2026-10-25
- Backdoored Models in Coding Agents: Provenance and Egress newly.app · Blog · 2026-10-25
- ProjectDiscovery Demonstrates Open-Weight Model Backdoor Risk (LetsDataScience) letsdatascience.com · News · 2026-10-06
- ProjectDiscovery built a sub-$50 model backdoor that got Codex to run a payload (RuntimeWire) runtimewire.com · News · 2026-10-25
Further reading
Drafted by the Threadlinqs research pipeline from 8 sources; published ; last updated —.