AI Briefing — June 6, 2026

🔬 Research

Code2LoRA: Hypernetwork-Generated Adapters for Code Language Models under Software Evolution

56 upvotes · Hotsko et al. · Paper

A hypernetwork framework that generates repository-specific LoRA adapters for code models, eliminating the need for expensive per-repository fine-tuning or long context injection via RAG. Code2LoRA-Static handles stable codebases while Code2LoRA-Evo maintains adapters that update with code diffs using GRU hidden states. On RepoPeftBench (604 Python repos), it achieves 63.8% cross-repo exact match, matching per-repository LoRA performance but with zero inference-time token overhead.

ArcANE: Do Role-Playing Language Agents Stay in Character at the Right Time?

42 upvotes · Song et al. · Paper

A benchmark testing whether role-playing language agents maintain character consistency as psychological arcs evolve across narrative phases, not just fixed persona recall. Spanning 17 novels and 80 characters, ArcANE evaluates responses both within source text and in novel scenarios the original never explored. Character Arc conditioning outperforms all other context strategies across six models, with the largest gains on out-of-source scenarios where retrieval methods fail.

TIDE: Proactive Multi-Problem Discovery via Template-Guided Iteration

36 upvotes · Jeong et al. · Paper

A framework for agents to proactively discover multiple hidden problems in user contexts beyond explicit requests, addressing the limitation that users only surface problems they've noticed. TIDE uses iterative discovery (small batches per round, conditioning on previous finds) and thought templates (reusable schemas from solved cases) to achieve better task coverage. Validated on personal workspaces and software repositories, it substantially outperforms single-shot and parallel multi-agent baselines.

🛠 Tools

chopratejas/headroom — Context compression layer for AI agents

2473 stars/day · Repo

A context compression system that reduces tokens by 60-95% while maintaining answer quality, targeting tool outputs, logs, files, and RAG chunks before they reach LLMs. Offers six compression algorithms and works as a library, proxy, or MCP server with local-first, reversible compression. With 15K GitHub stars and active development, it addresses the critical token cost problem facing AI agents and RAG systems.

NousResearch/hermes-agent — Self-improving AI agent with built-in learning

1845 stars/day · Repo

An AI agent that builds skills from experience, improves them during use, and creates a deepening model of users across sessions. Features cloud deployment ($5 VPS to GPU clusters), model flexibility (200+ via OpenRouter, Nous Portal, local endpoints), and cross-platform access including Telegram integration. With 184K stars, it represents a major push toward persistent, learning-enabled AI agents.

affaan-m/ECC — Agent harness performance optimization system

1361 stars/day · Repo

A performance optimization system for AI coding agents across Claude Code, Codex, Cursor, and similar platforms, focusing on skills, memory, security, and research-first development. With 209K stars and multilingual support, it positions itself as harness-native infrastructure for agentic workflows. The project emphasizes security through AgentShield components and offers GitHub App integration.

📡 Analysis

Running Python code in a sandbox with MicroPython and WASM

June 6 · Post · tags: python, sandboxing, ai, datasette, webassembly

Simon Willison details his latest approach to code sandboxing using MicroPython compiled to WebAssembly, addressing the need to run plugin code safely within Python applications. The solution offers memory/CPU limits, controlled file access, no network access, and clean PyPI installation—solving longstanding challenges for systems like Datasette and LLM plugins. This represents a practical breakthrough for AI systems that need to execute user-generated code safely.

OpenAI Help: Lockdown Mode

June 5 · Post · tags: security, ai, openai, prompt-injection

OpenAI has launched Lockdown Mode to prevent data exfiltration in the final stage of prompt injection attacks by limiting outbound network requests. Willison notes this directly attacks the "Lethal Trifecta" (private data + untrusted content + exfiltration vectors) by cutting off the easiest leg to restrict. The feature's existence implies ChatGPT's default settings don't provide robust protection against determined data exfiltration attacks.

Sources: HuggingFace Papers API, GitHub Trending, simonwillison.net