Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents
Source: Microsoft Research (opens in a new tab) · Dimitris Papailiopoulos
Intel Summary
Microsoft Research has identified "Lazarus," a backdoor technique that repurposes dormant neurons within an AI model to execute malicious instructions when encountering a hidden trigger. According to the researchers, the vector presents heightened risks for AI agents possessing terminal or external tool access, where a triggered model can covertly download and execute attacker code.
Why It Matters
The finding exposes a structural vulnerability where latent model parameters can be weaponized without disrupting baseline behavior. Organizations deploying autonomous AI agents with execution privileges or shell access face potential supply-chain and runtime risks, reinforcing the necessity for strict capability boundaries and sandboxing around agentic tool use.
Organizations & Entities
Topics
Related Intelligence
- DevelopmentNewAlso involving Microsoft Research
Microsoft Research identifies Lazarus backdoor technique in AI models
Microsoft Research has identified "Lazarus," a backdoor technique that repurposes dormant neurons within an AI model to execute malicious instructions when encountering a hidden trigger. Claims are as reported; this summary makes no determination about accuracy or significance.
- ReportRelated through Microsoft
Microsoft Labors to Fix Long Outage of Office, Outlook, Copilot
The Information reports that Microsoft experienced a global outage lasting over six hours, taking Microsoft Office, Outlook, and Copilot offline. The company stated it was actively attempting to restore services for affected users worldwide.
The Information - ReportRelated through Microsoft
Nebius tests the neocloud playbook with $40 billion in planned spending
Nebius plans to fund a $40 billion AI infrastructure expansion using customer prepayments and loans. The strategy faces scrutiny as delays at a Microsoft data center raise questions about the resilience of this capital-intensive financing model.
TechCrunch - ReportRelated through Microsoft
AI Inference Startups Are Quietly Buying Their Own Chips
AI inference startups that previously relied on leasing and reselling cloud capacity are beginning to purchase their own AI chips due to compute shortages. As major hyperscalers face tight capacity constraints, these providers are altering their capital structure to own physical hardware directly while increasingly turning to smaller neocloud providers for available compute resources.
The Information