Poisoned GGUF Templates Enable Inference-Time Supply Chain Attacks
GGUF
Researchers demonstrated an inference-time supply-chain backdoor in GGUF templates that alters model and agent behavior without changing weights. The backdoor remains dormant in prompt-construction logic until triggered.
- first report
- 31 Jul 2026
- last update
- 31 Jul 2026
Timeline · 3
Sources · 3
News and write-ups
- 2026-07-31LLM Backdoors at the Inference Level: The Threat of Poisoned TemplatesPillar Security · pillar.security/blog/llm-backdoors-at-the-inference-level-the-threat-of-poisoned
Research and catalogues
- 2026-07-31Poisoned GGUF Templates: Inference-Time Supply Chain AttackMITRE ATLAS · atlas.mitre.org/studies/AML.CS0064
- 2026-07-31Inference-Time Backdoors via Chat Templates: From LLM Supply Chains to Agentic System CompromisearXiv · arxiv.org/abs/2602.04653
Titles and links belong to their publishers. The summary above is written by gcdRadar in its own words; data is CC BY-SA 4.0.
Classification
- lane
- Research
- AI role
- Attack on AI
- categories
- Supply chainAgent hijack
- severity
- 1 · Research / proof of concept
- targets
- Open-weight model, Framework / tooling
- owasp llm top 10
- LLM03:2025 Supply ChainLLM06:2025 Excessive Agency
- dates
- happened Jun 2025 (approx.) · first reported 31 Jul 2026
- origin
- MITRE ATLASDerived from MITRE ATLAS (Apache 2.0, © MITRE).
- record
- confidence 0.95 · created 2026-10-18 06:55Z · updated 2026-10-18 06:55Z
Lane, AI role, categories and severity come from the classifier; the OWASP and ATLAS rows are derived from the categories. How these are assigned
Copy a brief
This record as a short Markdown brief, for a ticket, a chat or a coding agent.
the brief as text
# Poisoned GGUF Templates Enable Inference-Time Supply Chain Attacks https://gcdradar.vercel.app/incidents/poisoned-gguf-templates-enable-inference-time-supply-chain-a · AIR-2026-0359 · closed 2026-08-14 - Lane: Research - Severity: 1 (Research / proof of concept) - Fix status: Fix status unknown - First reported: 2026-07-31 (happened 2025-06) - AI role: Attack on AI - Products: GGUF ## Summary Researchers demonstrated an inference-time supply-chain backdoor in GGUF templates that alters model and agent behavior without changing weights. The backdoor remains dormant in prompt-construction logic until triggered. ## What to do No action recorded; read the sources below and the vendor's advisory. ## Latest updates - 2026-07-31: Inference-Time Backdoors via Chat Templates: From LLM Supply Chains to Agentic System Compromise (https://arxiv.org/abs/2602.04653) - 2026-07-31: LLM Backdoors at the Inference Level: The Threat of Poisoned Templates (https://pillar.security/blog/llm-backdoors-at-the-inference-level-the-threat-of-poisoned-templates) - 2026-07-31: Poisoned GGUF Templates: Inference-Time Supply Chain Attack (https://atlas.mitre.org/studies/AML.CS0064) ## Sources - 2026-07-31 MITRE ATLAS: Poisoned GGUF Templates: Inference-Time Supply Chain Attack https://atlas.mitre.org/studies/AML.CS0064 - 2026-07-31 Pillar Security: LLM Backdoors at the Inference Level: The Threat of Poisoned Templates https://pillar.security/blog/llm-backdoors-at-the-inference-level-the-threat-of-poisoned-templates - 2026-07-31 arXiv: Inference-Time Backdoors via Chat Templates: From LLM Supply Chains to Agentic System Compromise https://arxiv.org/abs/2602.04653 Verify against the vendor's advisory before changing production. Data CC BY-SA 4.0, https://gcdradar.vercel.app.
Related
- closedSteering Vectors Steering for Safety, Hijacking for Profit: Trojan Threats in LLM Steering
Research · since 27 Oct 2026
AI - closedSEAL: Persistent Tool Poisoning Attack against LLM Agents
Research · since 20 Oct 2026
AI - liveAI agents Microsoft Researchers Show AI Agents Can Be Backdoored via Neuron Edits
Research · since 11 Nov 2026 · updated 13 Nov 2026
AI - liveOpenAI Codex OpenAI Codex Jailbreak Toolkit Hides Prompt Adware to Hijack AI Agents
Research · since 4 Nov 2026
AI - liveLLM Agent Frameworks Beyond Prompt Injection: The Tool Update Gap in LLM Agent Security
Research · since 3 Nov 2026
AI - liveKerkoporta: Injecting Structured-Output Backdoors in Reasoning Coding Agents
Research · since 3 Nov 2026
AI