Executive Summary
Researchers have demonstrated a technique dubbed 'Lazarus' — described by Microsoft as a 'sleeper agent' attack — that embeds malicious payloads inside the weights of open-weight AI models using fewer than ten dormant neurons. Because open-weight models consist of billions of parameters, a payload occupying as little as 0.00001% of a neural network is virtually undetectable through normal inspection or behavioral testing. Once a predefined trigger — such as a specific user input, tool output, or even a system date — fires the dormant neurons, the encoded command executes within the model's runtime environment, potentially granting attackers access to credentials, code, and connected systems.
This technique represents an escalation in AI supply-chain risk. As 63% of organizations already use open-weight models and adoption is accelerating, the attack surface is expanding rapidly. Unlike traditional software vulnerabilities, model-weight poisoning leaves no readable code for security scanners to flag, making conventional defenses insufficient and elevating the urgency for network-layer controls around AI agent environments.
Why This Matters Now
Open-weight AI model adoption surged past 63% of enterprises by early 2025 and continues to grow. With platforms like Hugging Face hosting hundreds of thousands of models — and malicious models already discovered there exploiting Pickle serialization — the shift to weight-embedded payloads marks a stealthier, harder-to-detect evolution. Regulatory frameworks have not yet caught up, and most organizations lack tooling to diff model weights against trusted baselines, leaving AI agent pipelines exposed to undetected, trigger-activated compromise at scale.
Attack Path Analysis
An attacker poisons an open-weight AI model by encoding a malicious payload across dormant neurons before the model is published to a public repository such as Hugging Face; a victim organization downloads and deploys the model, granting it access to code, credentials, and internal systems. Once a trigger condition is met, the sleeper agent activates, escalates access using available credentials, moves laterally across workloads, establishes covert command-and-control via unexpected network connections, exfiltrates sensitive data through permitted egress paths, and ultimately disrupts or compromises production systems.
Kill Chain Progression
This analysis maps confirmed threat intelligence to the full cloud kill chain to show where defensive gaps would emerge as an attack progresses.
Initial Compromise
Description
Attacker publishes a subtly poisoned open-weight AI model to a public model repository. The model's weights encode a malicious payload across dormant neurons using techniques analogous to MaleficNet or the Microsoft-described 'Lazarus' method. A developer or organization downloads the model and integrates it into their environment, trusting its apparent legitimacy. No CVE is required; the attack abuses the trust model inherent in open-weight AI distribution, similar to supply chain compromise via open-source packages.
MITRE ATT&CK® Techniques
Supply Chain Compromise: Compromise Software Dependencies and Development Tools
Compromise Host Software Binary
Obfuscated Files or Information
Command and Scripting Interpreter
User Execution: Malicious File
Unsecured Credentials
Escape to Host
Application Layer Protocol
Potential Compliance Exposure
Mapping incident impact across multiple compliance frameworks.
NIST SP 800-53 Rev 5 – Provenance and Integrity of Third-Party Components
Control ID: SR-4
CISA Zero Trust Maturity Model 2.0 – AI Agent Least-Privilege Access Enforcement
Control ID: Applications Pillar – Authorization and Least Privilege
NIS2 Directive – Supply Chain Security and Secure Software Acquisition
Control ID: Article 21(2)(a) and 21(2)(d)
DORA (EU 2022/2554) – Third-Party AI Component Risk Assessment
Control ID: Article 28 – ICT Third-Party Risk Management
PCI DSS 4.0 – Secure Development and Third-Party Software Risk Management
Control ID: Requirement 6.3 and 12.3.4
ISO/IEC 27001:2022 – AI Supply Chain Integrity and Supplier Security Controls
Control ID: Annex A 8.30 – Outsourced Development and A 5.19 – Information Security in Supplier Relationships
Sector Implications
Industry-specific impact of the vulnerabilities, including operational, regulatory, and cloud security risks.
Information Technology/IT
AI supply-chain poisoning via open-weight model tampering exposes IT environments to sleeper-agent payloads; zero trust segmentation and CNSF controls are critical to contain compromised model lateral movement and credential exfiltration.
Computer Software/Engineering
Software teams downloading open-weight models from Hugging Face face model poisoning risks; dormant neuron payloads can exfiltrate credentials and code, demanding egress policy enforcement and multicloud visibility aligned to NIST.800-53.SC-18.
Financial Services
AI-driven financial workflows adopting open-weight models risk sleeper-agent compromise triggering unauthorized data exfiltration; PCI 4.0 and ZTMM compliance gaps in egress filtering and east-west traffic security amplify systemic exposure.
Health Care / Life Sciences
Healthcare AI deployments using poisoned open-weight models risk HIPAA-regulated data exfiltration via triggered dormant neurons; inline IPS, zero trust segmentation, and anomaly detection are essential mitigations per HIPAA.164.312 controls.
Sources
- 'Sleeper Agent' Tactic Hides Payloads in Open-Weight AI Modelshttps://www.darkreading.com/vulnerabilities-threats/sleeper-agent-tactic-payloads-open-weight-ai-modelsVerified
- MaleficNet: Hiding Malware in Neural Networkshttps://arxiv.org/abs/2209.11859Verified
- Hugging Face Security: Malicious Models and Pickle Exploitshttps://huggingface.co/blog/pickle-safetensorsVerified
- CISA Guidance on AI and Machine Learning Security Riskshttps://www.cisa.gov/topics/artificial-intelligenceVerified
- MITRE ATLAS: ML Supply Chain Compromise (AML.T0010)https://atlas.mitre.org/techniques/AML.T0010Verified
- Microsoft Security Blog: Defending Against Adversarial Machine Learning Threatshttps://www.microsoft.com/en-us/security/blog/Verified
Frequently Asked Questions
Cloud Native Security Fabric Mitigations and ControlsCNSF
Based on the attack progression modeled above, these are the defensive controls that would constrain each stage.
Aviatrix Zero Trust CNSF is relevant here because a poisoned AI model, once deployed and triggered, relies heavily on broad network reachability and permissive egress paths to escalate, pivot, and exfiltrate. By enforcing workload-level segmentation, identity-scoped east-west controls, and FQDN-aware egress policy, CNSF would likely reduce the blast radius available to the activated payload across each stage of this supply-chain-rooted attack.
Control: Cloud Native Security Fabric (CNSF)
Mitigation: CNSF fabric-level visibility and workload segmentation would likely limit the newly deployed AI model's immediate reachability to sensitive internal systems and credential stores, reducing the network surface available to the payload at the moment of deployment.
Control: Zero Trust Segmentation
Mitigation: Zero Trust segmentation would likely constrain the payload's ability to reach credential stores or escalate into adjacent service identities by enforcing identity-aware access boundaries around the AI workload's permitted service scope.
Control: East-West Traffic Security
Mitigation: East-west traffic enforcement would likely reduce the payload's lateral reachability by restricting service-to-service paths to only declared, policy-approved flows, making opportunistic pivots to adjacent workloads or data stores considerably harder to execute.
Control: Multicloud Visibility & Control
Mitigation: Multicloud visibility and centralized traffic control would likely surface unexpected outbound connection attempts from the AI workload and could reduce the payload's ability to establish reliable command-and-control channels by flagging anomalous destinations against baseline egress behavior.
Control: Egress Security & Policy Enforcement
Mitigation: FQDN-aware egress policy enforcement would likely reduce the payload's ability to exfiltrate data through permitted outbound paths by constraining outbound traffic to explicitly approved external destinations and flagging or blocking connections to unknown endpoints.
Because CNSF controls would likely have reduced lateral reachability, constrained egress paths, and limited the credential and service scope available to the payload across earlier stages, any residual impact would more likely be scoped to the AI workload's immediate environment rather than spreading across production systems, cloud accounts, or downstream business processes.
Impact at a Glance
Affected Business Functions
- AI Model Procurement and Deployment Pipelines
- Software Development and AI-Assisted Coding Environments
- Credential and Secrets Management Systems
- Enterprise AI Agent Orchestration and Automation Workflows
Estimated downtime: N/A
Estimated loss: N/A
This technique poses a latent, high-severity risk to any organization deploying open-weight AI models from public repositories such as Hugging Face. If a poisoned model is deployed, adversaries could exfiltrate credentials, API keys, source code, and sensitive system data accessible to the AI agent at runtime. The dormant payload is designed to remain inactive until a configured trigger is met, meaning data exposure could occur silently over an extended period without detection. Categories of data at risk include: developer credentials and secrets stored in environment variables, proprietary source code processed by AI coding assistants, internal system architecture details, and any production data accessible to AI agents granted broad permissions. The financial and reputational impact of a successful exploitation at an enterprise scale could be significant, though no confirmed real-world exploitation of weight-embedded payloads has been publicly documented as of this report.
Recommended Actions
Key Takeaways & Next Steps
- • Apply Cloud Native Security Fabric (CNSF) inline enforcement at the AI agent's network boundary to detect and block anomalous outbound connections initiated by triggered payloads, treating AI model runtimes as untrusted workloads subject to real-time inspection and policy enforcement.
- • Enforce Zero Trust Segmentation with least-privilege identity-based policies for all AI agents: restrict service identities to only the APIs, credentials, and namespaces required for their defined task scope, so that a compromised model cannot escalate or move laterally beyond its minimal footprint.
- • Deploy Egress Security and Policy Enforcement with FQDN filtering on all AI agent outbound traffic to block connections to unauthorized destinations; require explicit allowlisting of external endpoints and flag or block any novel outbound domain or IP not in the approved egress policy.
- • Enable Multicloud Visibility and Control to baseline normal AI agent behavior—including expected tool invocations, API call patterns, and network destinations—and alert on deviations such as unexpected shell commands, credential access outside normal context, or invocation of tools in unusual sequences.
- • Use East-West Traffic Security controls to monitor and restrict workload-to-workload communication initiated by AI agents; ensure that inter-service calls from model runtimes are logged, inspected, and subject to microsegmentation policy so lateral movement through legitimate agent communication paths is detected and contained.



