🧬 UnifyBio: Power Tools for Translational Data Science
🔗 Read more 🔗
🛡️ Task Injection: Attacking Autonomous AI Agents
As agents become more capable, this work maps a critical new attack surface and reinforces the need for layered AI safety strategies.
Google’s security research team describes “Task Injection” attacks that target autonomous AI agents by abusing seemingly legitimate sub-tasks to trigger harmful behavior or data leaks. Unlike classic prompt injection, these attacks exploit agent autonomy. The post analyzes vulnerabilities discovered in OpenAI’s Operator and outlines mitigations such as runtime policy enforcement and human oversight.
🔗 Read more 🔗
🔍 Text Similarity via Normalized Compression Distance
A thoughtful blend of classical algorithms and AI-assisted coding, highlighting both where LLMs help and where they still fall short.
The author describes building a Haskell library for text similarity search based on Normalized Compression Distance. They discuss challenges with LLM-generated Haskell code, debugging type errors, and improving correctness through property-based testing. The post also shows how compression-based methods can act as a viable alternative to neural embeddings.
🔗 Read more 🔗