LLMs Can Transmit Hidden Traits Through Unrelated Data, Study Finds
AI researchers demonstrated that a teacher model's hidden preferences—such as favoring owls—can be passed to a student model via training data completely unrelated to that trait, like sequential numbers. Even explicit filtering fails to block this "subliminal learning" effect; in one test, a GPT-4.1 nano student's owl-preference jumped from 12% to over 60% after fine-tuning on number sequences generated by the biased teacher.
In security news, operatives from the Lazarus Group were lured into a fake DeFi startup, Ballena Azul LTD, where sandboxed laptops captured their every action, revealing remote access tools and AI usage. Separately, researchers found that malicious SIM cards can command phones to leak data, force 2G downgrades, or execute code, with vulnerabilities confirmed across 26 devices. Another team discovered a method to extract hidden reasoning traces from proprietary LLM APIs, raising concerns about data leaks and distillation attacks.
On the infrastructure side, a new Metal capability shim from Cua boosted LLM inference on Apple Silicon macOS VMs by 11–16x in llama.cpp by correcting outdated GPU capability reporting. Mistral AI unveiled plans to build 1 gigawatt of European compute by 2030 (200 MW by 2027) and introduced a Priority Tier with uptime guarantees for mission-critical AI workloads.
Open model advances include a collaboration between Loka and Arcee AI to train a scientific investigation model that uses tools and reasoning to infer Gene Ontology annotations with higher accuracy. A series on multi-tenant memory design for AI agents explored tiered recall across short-, medium-, and long-term memory, using Alice and Bob scenarios to illustrate scoping choices.
In developer tools and practices, Stock+ migrated from Firebase Cloud Messaging to OneSignal, adopting a durable inbox and identity-based notifications to cut complexity. And a deep dive into Linux capabilities and seccomp filters explained how to effectively layer privileges and system call restrictions, using CAP_AUDIT_WRITE as a practical example.
» Statistics
- Posts
- 27
- Reads
- 1
- Avg. score
- 7.5
» Most read
- We Reduced Our AWS Bill by 10× by Checking It Thoroughly
- Previous-Token Prediction Based LLM Near-Exact Prompt Reconstruction
- Open-Source eBPF Volumetric DDoS Protection
- Security Mining Report: Challenges in Val Town
- Supply Chain Attack Discovered in BdThemes WordPress Plugins
- Shipping OneSignal for Shipaton
- When Configuration Management Becomes an Operational Liability
- Harvesting Ethereum Traces Without an Archive Node
- Apple Silicon macOS VMs Get 11-16x Faster LLM Inference via Llama.cpp
- No Standard Defines What to Record for an LLM Call, So I Built One
» Top scored
- Study: LLMs Can Transmit Hidden Traits Through Unrelated Data
- Apple Silicon macOS VMs Get 11-16x Faster LLM Inference via Llama.cpp
- Researchers Lure Lazarus's Fake IT Workers Into a Sham DeFi Startup
- Stealing Reasoning Traces from Proprietary LLM APIs
- Open Model Enhanced for Scientific Inquiry
- Defense-in-Depth and Linux Capabilities: A Study on Seccomp Filters
- Shipping OneSignal for Shipaton
- Malicious SIM Cards Can Shut Down Phones and Steal Files
- Mistral AI aims to build 1 gigawatt of European compute by 2030
- Building Multi-Tenant, Multi-Tier Memory for AI Agents