[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$fBaUWG5BjMjlOu9KM3q6vdAbNEUd8Dqcdn9V4iDdJiTA":3},{"lesson":4},{"id":5,"slug":6,"article_id":7,"title":8,"body":9,"prevention":10,"framework_refs":11,"framework_notes":17,"status":23,"created_at":24,"published_at":25,"article":26,"tags":30,"podcasts":43,"controls":44},"df3a15f0-8b6d-4c95-ae6e-220489769d5d","hidden-ai-backdoors-can-survive-standard-evaluations-via-dormant-neuron-manipulation","166b51e1-619e-44d4-b7da-0fe68e0c6491","Hidden AI Backdoors Can Survive Standard Evaluations via Dormant Neuron Manipulation","Microsoft researchers discovered that attackers can embed hidden backdoors inside AI model weights by manipulating dormant neurons — neurons that are normally inactive but can be activated by a specific attacker-controlled trigger. Once triggered, the compromised AI agent can download and execute malicious code without any external malware loader being present. Because the backdoor lives entirely within the model's internal parameters, standard safety evaluations and behavioral testing fail to detect it. This matters because organizations increasingly deploy third-party or open-source AI models in production pipelines, creating a new and largely unmonitored supply chain attack surface. Trust in a model's benchmark performance is no longer sufficient assurance of its integrity.","**Immediate actions:**\n- Audit all AI models in production or pre-production to document their provenance, training data sources, and fine-tuning history.\n- Restrict AI agents' runtime permissions so they cannot make outbound network calls or execute downloaded code without explicit authorization.\n\n**Long-term improvements:**\n- Establish a formal AI model vetting process that includes adversarial red-teaming and trigger-probing tests beyond standard benchmarks before deployment.\n- Treat AI model files as untrusted third-party software and apply supply chain controls — cryptographic signing, hash verification, and sourcing only from trusted repositories.\n- Develop and maintain an inventory of all deployed AI models with version tracking so compromised versions can be rapidly identified and replaced.\n\n**Detection measures:**\n- Implement runtime behavioral monitoring for AI agents to flag anomalous actions such as unexpected network connections or code execution attempts.\n- Log all AI agent inputs and outputs in a tamper-evident store to enable forensic reconstruction if a trigger event is suspected.",[12,13,14,15,16],"MITRE ATT&CK T1195: Supply Chain Compromise","MITRE ATT&CK T1059: Command and Scripting Interpreter","NIS2 Article 21(2)(d)","NIS2 Article 21(2)(f)","NIST SP 800-53 SA-12: Supply Chain Protection",{"NIS2 Article 21(2)(d)":18,"NIS2 Article 21(2)(f)":19,"MITRE ATT&CK T1195: Supply Chain Compromise":20,"NIST SP 800-53 SA-12: Supply Chain Protection":21,"MITRE ATT&CK T1059: Command and Scripting Interpreter":22},"NIS2 requires security in supply chain relationships, and this incident shows that AI models sourced externally must be subject to integrity verification and vetting as part of that obligation.","This article requires organizations to assess the effectiveness of their security measures, and standard model evaluations were shown here to be insufficient to detect the hidden backdoor, demanding more rigorous assurance methods.","The Lazarus technique embeds malicious functionality into the AI model artifact itself during or after training, mirroring supply chain compromise where a trusted component is weaponized before delivery to the victim organization.","SA-12 requires organizations to protect against supply chain threats to systems and components, which directly applies to ensuring AI model weights have not been tampered with before or during deployment.","Once the dormant neuron backdoor is triggered, the AI agent downloads and executes malicious code, which maps directly to adversary use of execution techniques through a trusted process.","published","2026-11-11T12:20:24.19046+00:00","2026-11-11T12:20:24.174+00:00",{"id":7,"url":27,"slug":28,"title":29},"https:\u002F\u002Fcybersecuritynews.com\u002Fclaude-code-ai\u002F","claude-code-ai-discovers-technique-to-turn-dormant-model-neurons-into-hidden-mal-d3c5e2","Claude Code AI Discovers Technique to Turn Dormant Model Neurons Into Hidden Malware Triggers",[31,37],{"id":32,"name":33,"slug":34,"description":35,"color":36},"05757c8d-6b93-4194-b35d-7359e7d33b0e","Vulnerability Management","vulnerability-management","Missing scans, no risk prioritization","#fb923c",{"id":38,"name":39,"slug":40,"description":41,"color":42},"f0c2a0af-58aa-4128-87c9-6acd30f2dc48","Supply Chain","supply-chain","Third-party risk, compromised dependencies","#8b5cf6",[],[45,48,52],{"slug":34,"name":46,"problem":47},"Vulnerability scanning and prioritisation","You know which systems are vulnerable, and which ones to fix first.",{"slug":49,"name":50,"problem":51},"endpoint-detection-response","Endpoint detection and response (EDR and XDR)","Malware and hands-on-keyboard attackers on laptops and servers get spotted and stopped.",{"slug":53,"name":54,"problem":55},"ai-security","AI application security","Prompt injection, data leaks through AI tools and unapproved AI use get caught."]