groundy

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

agents

When MCP Tool Descriptions Don't Match the Code, Agents Trust the Lie

A study of 2,214 MCP servers finds 9.93% of tool descriptions diverge from the code, creating a confused-deputy risk for agent runtimes that select tools by description alone.

6min · 19:19 · jun 03
security

Students Are Prompt-Injecting AI Graders to Score Full Marks

A June 2026 arXiv study finds that prompt injection in student submissions manipulates LLM grading systems into awarding full marks, and current defenses do not hold.

6min · 16:22 · jun 03
devtools

Malicious npm Packages Hit Red Hat's Published JavaScript Clients

Malicious versions of 32 Red Hat npm packages carried a credential-stealing worm, published through the vendor's OIDC pipeline. Vendor namespaces are not a trust boundary.

policy

Stacked Org Policies in LLM Chatbots Break Where Rules Collide

Stacking HR, legal, and brand policies in LLM prompts assumes additive compliance. Graph-based research finds per-rule testing misses combinatorial policy conflicts.

security

Removing an LLM Backdoor Post-Training Without the Poisoned Data

Patcher removes LLM backdoor triggers from a single observed failure and model weights, no poisoned training data required. Deployers gain an alternative to full retraining.

models

Which Layer Detects LLM Hallucinations Best? The Case Against Fixed-Layer Probes

An ICML 2026 paper finds that fixed-layer hallucination probes miss detection signal, and proposes FEPoID, a training-free method to calibrate layer choice per model.

policy

Why Fine-Tuning Strips Safety Alignment From Open-Weight LLMs

Safety alignment in open-weight LLMs is concentrated in a handful of output tokens. Benign fine-tuning erases them, making release-time safety evaluations unreliable.

security

Stored Prompt Injection Now Persists Across AI Agent Sessions

Prompt injection planted in one agent session resurfaces in later ones through persistent memory and tool state, bypassing input sanitization that only validates external.

featured editor's picks
most popular last 6 days · by traffic
top highest rated
feed all 370 articles · reverse chronological
  1. jun 03 agents When MCP Tool Descriptions Don't Match the Code, Agents Trust the Lie
  2. jun 03 security Students Are Prompt-Injecting AI Graders to Score Full Marks
  3. jun 03 devtools Malicious npm Packages Hit Red Hat's Published JavaScript Clients
  4. jun 03 policy Stacked Org Policies in LLM Chatbots Break Where Rules Collide
  5. jun 03 security Removing an LLM Backdoor Post-Training Without the Poisoned Data
  6. jun 03 models Which Layer Detects LLM Hallucinations Best? The Case Against Fixed-Layer Probes
  7. jun 03 policy Why Fine-Tuning Strips Safety Alignment From Open-Weight LLMs
  8. jun 03 security Stored Prompt Injection Now Persists Across AI Agent Sessions
  9. jun 03 industry MiniMax M3 Bundles 1M Context and Native Multimodal Into One Open-Weight Model
  10. jun 03 security LLM Data Poisoning Survives the Data-Cleaning Defenses Built to Stop It
  11. jun 03 devtools OpenAI Upgrades Codex Right as Teams Weigh Leaving Claude Code
  12. jun 03 policy Game Theory vs RLHF: Modeling LLM Safety Alignment as a Non-Cooperative Game
  13. jun 03 infra Cost-Aware RAG Routing: When Deeper Retrieval Stops Paying Off
  14. jun 02 devtools GitHub Copilot Moves to a Platform App, Decoupling From the Editor
  15. jun 02 infra Using Your Nvidia GPU's VRAM as Linux Swap: Where the NBD Hack Breaks Down
  16. jun 02 security Why OpenAI Bets on Instruction Hierarchy to Stop Prompt Injection
  17. jun 02 policy Explainability Mandates Leak Graph Models to Their Attackers
  18. jun 02 security Stopping Multi-Turn LLM Jailbreaks Without Retraining the Model
  19. jun 02 security African Languages Are a Jailbreak Blind Spot for English-Tuned LLM Safety
  20. jun 02 devtools How a VSCode Bug Let One Click Steal Your GitHub Token
  21. jun 02 agents When an AI Agent Causes a Loss, Who Files the Insurance Claim?
  22. jun 02 models Cross-Domain RL Training Degrades Capabilities. CARE-RL Reweights to Fix It
  23. jun 02 agents When Agent Skill Libraries Scale, Dependency-Aware Retrieval Beats Flat Search
  24. jun 02 policy Evolutionary Search Finds LLM Jailbreak Classes That Static Red-Teaming Misses
  25. jun 02 security Poisoning Open-Source LLM Merges: One Bad Checkpoint Hijacks the Result
  26. jun 02 agents Can Instruction-Tuned Retrievers Fix Agentic Search's Retrieval Gap?
  27. jun 02 models LLM Watermarking Without Quality Loss: The Non-Distortionary Approach
  28. jun 02 security An Autonomous Research Agent Now Discovers SOTA LLM Jailbreak Attacks
  29. jun 02 devtools GitHub Copilot and Productivity: What an Observational Dose-Response Study Measures
  30. jun 02 policy Why AI Red-Teaming Rediscovers the Same Jailbreaks and Misses the Rest
  31. jun 02 industry Morningstar's $780B SpaceX Mark Undercuts the IPO Target by Half
  32. jun 02 security Malware Can Prompt-Inject the AI Agent Reverse-Engineering It
  33. jun 02 agents Bandit-Based Prompt Optimization Targets Multi-Agent Systems Like CrewAI and AutoGen
  34. jun 01 oss Open-Source Workspace Suite tinycld Takes On Google and Nextcloud
  35. jun 01 oss DARPA's AIxCC Postmortem: What Autonomous Cyber Reasoning Systems Got Right and Wrong
  36. jun 02 security CVE-Factory Turns Published CVEs Into Security Agent Training Data. A 32B Model Beats Claude 4.5 Sonnet.
  37. jun 01 oss An Open-Source Home Camera That Encrypts End-to-End Instead of Trusting Ring
  38. jun 01 policy LLMs Treat the Assistant Persona as Privileged. That's a Safety Gap
  39. jun 01 industry Vercel's Grep Buy Signals Code Search Is Now AI Agent Infrastructure
  40. jun 01 security LLM Reasoning Traces Leak the Private Data They're Told to Hide
  41. jun 01 models Treating LLM Agent Memory as a Database: The VikingMem Approach
  42. jun 01 oss Your Open-Source License Won't Stop Someone Phishing With Your Code
  43. jun 01 models Can a Language Model Work Without a Neural Network? A New arXiv Paper Says Yes
  44. jun 01 models Can Code-Generating LLMs Do Engineering Math? FEM-Bench Tests Them
  45. jun 01 policy Newer LLMs Aren't Always Safer: Adversarial Attacks Transfer Across Model Generations
  46. jun 01 models Unlearning Isn't Deletion: arXiv 2505.16831 Shows Machine Unlearning in LLMs Is Reversible
  47. jun 01 security Video Jailbreaks Hit Multimodal LLMs by Splitting Payloads Across Clips
  48. jun 01 industry OMB's Power to Cancel Any Grant at Any Time Shifts Risk Onto University AI Labs
  49. may 31 models Why LLMs Fail at Spatial Reasoning When Planning Navigation
  50. may 31 culture Ranking LLMs Side by Side Makes Their Dialect Bias Worse
load older →