groundy

Groundy — independent coverage of developer tools, infrastructure, and platforms

models

Can You Stitch Two Foundation Models Together Without Retraining?

Splicing layers from independently trained foundation models fails without targeted training at the join point. A two-stage recipe called Final Feature Matching makes it work.

6min · 15:26 · jun 04
infra

Cloudflare Acquires VoidZero, the Company Behind Vite's Rust Toolchain

Cloudflare acquired VoidZero, putting Vite, Rolldown, and Oxc maintainers on a deploy-target vendor's payroll. MIT licensing stays. Roadmap neutrality is the open question.

6min · 12:08 · jun 04
security

Jailbreak Suffixes Hit Harder at Specific Token Positions, New GCG Variant Shows

SlotGCG shows adversarial token position, not just content, determines jailbreak success, with 14% higher attack rates and 42% higher rates against defended models.

policy

When Should an LLM Forget You? A Benchmark for Deciding What Memory to Drop

PersistBench finds LLMs mishandle persistent memory 53 to 97 percent of the time. Unlearning suppresses rather than erases user data, making GDPR compliance unverifiable.

security

OpenAI Adds Lockdown Mode to ChatGPT, Shifting Prompt-Injection Risk to Users

OpenAI's Lockdown Mode disables agentic features builders rely on rather than fixing prompt injection at runtime, forcing a binary choice between security and capability.

policy

When RL Training Rewards Capability-Seeking: A New Alignment Risk

A June 2026 ICML paper shows RL optimizers can push language models to exploit reward loopholes the task never required, while standard performance metrics hold steady.

models

Do Reasoning LLMs Waste Tokens? OckBench Tries to Measure It

OckBench scores 37 reasoning LLMs on token efficiency alongside accuracy, finding comparably accurate models differ by up to 26× in token cost under per-token billing.

security

Activation Steering Was Sold as LLM Control. New Work Makes It an Attack Surface

Poisoning 4-6% of tokens in a steering dataset silently inverts refusal vectors into jailbreaks, achieving 20-55% ASR. Shared vector bundles are the attack surface.

featured editor's picks
most popular last 6 days · by traffic
top highest rated
feed all 386 articles · reverse chronological
  1. jun 04 models Can You Stitch Two Foundation Models Together Without Retraining?
  2. jun 04 infra Cloudflare Acquires VoidZero, the Company Behind Vite's Rust Toolchain
  3. jun 04 security Jailbreak Suffixes Hit Harder at Specific Token Positions, New GCG Variant Shows
  4. jun 04 policy When Should an LLM Forget You? A Benchmark for Deciding What Memory to Drop
  5. jun 04 security OpenAI Adds Lockdown Mode to ChatGPT, Shifting Prompt-Injection Risk to Users
  6. jun 04 policy When RL Training Rewards Capability-Seeking: A New Alignment Risk
  7. jun 04 models Do Reasoning LLMs Waste Tokens? OckBench Tries to Measure It
  8. jun 04 security Activation Steering Was Sold as LLM Control. New Work Makes It an Attack Surface
  9. jun 04 culture Can Teaching Logical Fallacies Inoculate People Against AI Misinformation?
  10. jun 04 devtools Vercel Ships Experimental Native CLI Binaries to Cut the Node Startup Tax
  11. jun 04 security Catching LLM Agents Leaking Credentials From Their Own Activations
  12. jun 04 policy Refusal Steering Targets Individual Experts in MoE LLMs
  13. jun 04 infra Putting a Datacenter V100 in a Gaming PC: The Local LLM Math
  14. jun 04 devtools Vercel Rebuilds Its Marketplace CLI for Agents Instead of Humans
  15. jun 04 security The 2026 npm Attacks Proved AI Coding Assistants Are a Supply-Chain Target
  16. jun 03 security ChatGPT's New Lockdown Mode Borrows Apple's Name for a Prompt-Injection Kill Switch
  17. jun 03 agents When MCP Tool Descriptions Don't Match the Code, Agents Trust the Lie
  18. jun 03 security Students Are Prompt-Injecting AI Graders to Score Full Marks
  19. jun 03 devtools Malicious npm Packages Hit Red Hat's Published JavaScript Clients
  20. jun 03 policy Stacked Org Policies in LLM Chatbots Break Where Rules Collide
  21. jun 03 security Removing an LLM Backdoor Post-Training Without the Poisoned Data
  22. jun 03 models Which Layer Detects LLM Hallucinations Best? The Case Against Fixed-Layer Probes
  23. jun 03 policy Why Fine-Tuning Strips Safety Alignment From Open-Weight LLMs
  24. jun 03 security Stored Prompt Injection Now Persists Across AI Agent Sessions
  25. jun 03 industry MiniMax M3 Bundles 1M Context and Native Multimodal Into One Open-Weight Model
  26. jun 03 security LLM Data Poisoning Survives the Data-Cleaning Defenses Built to Stop It
  27. jun 03 devtools OpenAI Upgrades Codex Right as Teams Weigh Leaving Claude Code
  28. jun 03 policy Game Theory vs RLHF: Modeling LLM Safety Alignment as a Non-Cooperative Game
  29. jun 03 infra Cost-Aware RAG Routing: When Deeper Retrieval Stops Paying Off
  30. jun 02 devtools GitHub Copilot Moves to a Platform App, Decoupling From the Editor
  31. jun 02 infra Using Your Nvidia GPU's VRAM as Linux Swap: Where the NBD Hack Breaks Down
  32. jun 02 security Why OpenAI Bets on Instruction Hierarchy to Stop Prompt Injection
  33. jun 02 policy Explainability Mandates Leak Graph Models to Their Attackers
  34. jun 02 security Stopping Multi-Turn LLM Jailbreaks Without Retraining the Model
  35. jun 02 security African Languages Are a Jailbreak Blind Spot for English-Tuned LLM Safety
  36. jun 02 devtools How a VSCode Bug Let One Click Steal Your GitHub Token
  37. jun 02 agents When an AI Agent Causes a Loss, Who Files the Insurance Claim?
  38. jun 02 models Cross-Domain RL Training Degrades Capabilities. CARE-RL Reweights to Fix It
  39. jun 02 agents When Agent Skill Libraries Scale, Dependency-Aware Retrieval Beats Flat Search
  40. jun 02 policy Evolutionary Search Finds LLM Jailbreak Classes That Static Red-Teaming Misses
  41. jun 02 security Poisoning Open-Source LLM Merges: One Bad Checkpoint Hijacks the Result
  42. jun 02 agents Can Instruction-Tuned Retrievers Fix Agentic Search's Retrieval Gap?
  43. jun 02 models LLM Watermarking Without Quality Loss: The Non-Distortionary Approach
  44. jun 02 security An Autonomous Research Agent Now Discovers SOTA LLM Jailbreak Attacks
  45. jun 02 devtools GitHub Copilot and Productivity: What an Observational Dose-Response Study Measures
  46. jun 02 policy Why AI Red-Teaming Rediscovers the Same Jailbreaks and Misses the Rest
  47. jun 02 industry Morningstar's $780B SpaceX Mark Undercuts the IPO Target by Half
  48. jun 02 security Malware Can Prompt-Inject the AI Agent Reverse-Engineering It
  49. jun 02 agents Bandit-Based Prompt Optimization Targets Multi-Agent Systems Like CrewAI and AutoGen
  50. jun 02 security CVE-Factory Turns Published CVEs Into Security Agent Training Data. A 32B Model Beats Claude 4.5 Sonnet.
load older →