Security
Can Provable Bounds Defend LLM Fine-Tuning Against Poisoned Data?
BBoxER, a 2025 gradient-free post-training method, claims non-vacuous poisoning-robustness bounds for LLMs. The abstract never states how tight those bounds are.
The Groundy archive · Page 30 of 43
All articles
Browse Groundy's complete archive of 1027 articles on AI, developer tools and infrastructure. Page 30 of 43.
The archive
697–720 of 1027 articles · Newest first
- Developer Tools
Yarn Berry on Vercel: A Build-Cache Gap With No Documented Fix
Vercel's build cache excludes Yarn Berry's .yarn layout, so node-linker monorepos re-resolve dependencies every deploy. A 2024 YARN_CACHE_FOLDER workaround remains the fix.
- Infrastructure & Runtime
Turso on Vercel is a remote database integration, not embedded SQLite
Turso on Vercel is a remote database integration, not embedded SQLite inside each serverless function.
- Developer Tools
SvelteKit Can Run NextAuth.js, but Auth.js Moved to Better Auth
Auth.js v5 ships SvelteKit, SolidStart, Express, and Qwik bindings, freeing NextAuth.js from Next.js. But the project now lives under Better Auth, the new greenfield default.
- Agents & Frameworks
How On-Device AI Agents Keep Learning by Forgetting on Purpose
A frozen on-device agent learns new tasks by forgetting low-value memory on purpose, cutting its footprint 2.7x and prompt-injection success from 0.75 to zero.
- Developer Tools
Fired for Building the Google Workspace CLI: The Risk of Depending on Unofficial Vendor Tools
A Google engineer says he was fired for building gws, the unofficial Workspace CLI. With no SLA, teams that wired it into CI now own the migration.
- Models & Research
Flow Matching vs U-Net: A Skip-Free Backbone for Speech Models
A June 2026 preprint argues U-Net skip connections leak noise into flow-matching speech decoders, and a codec-supervised skip-free backbone can replace them at parity.
- Security
Measuring LLM Safety by Refusal Alignment Instead of Attack Success Rate
A June 2026 preprint proposes RAS, a white-box metric that scores LLM safety by hidden-state refusal alignment rather than blocked output, challenging ASR-only leaderboards.
- Security
Poisoning Physics-Informed Neural Networks Slips Past Loss-Based Validation
A June 2026 preprint shows poisoned physics-informed neural networks hit clean training loss while their solutions diverge up to 128%, defeating loss-based validation.
- Ethics, Policy & Safety
50 Years of Aviation Certification Expose a Structural Gap in AI Governance
A June 2026 arXiv paper argues AI governance lacks the epoch limits and proof surfaces that 50 years of aviation certification built in, breaking one-time approval stamps.
- Security
Catching LLM Jailbreaks by Watching Per-Layer Entropy, Not Outputs
A June 2026 paper reports jailbreaks perturb per-layer entropy of frozen LLMs before any harmful token emits, but adaptive attackers will likely follow one layer deeper.
- Open Source
DeepSeek, Qwen and GLM: Check license, access and serving needs
DeepSeek-V3's repository code is MIT-licensed; its Base and Chat weights have a separate Model License.
- Models & Research
A Per-Neuron Sequence Model Was Withdrawn From arXiv as Coverage Hailed It
TND proposed per-neuron dynamics as a sequence-modeling primitive, then was withdrawn from arXiv for accuracy errors the day before coverage called it a Transformer.
- Ethics, Policy & Safety
Do Reasoning Tokens Actually Make LLMs Safer? A New Paper Tests It
A June 2026 preprint finds refusal decisions are locked at a model's first token, undercutting the safety case for premium reasoning modes billed per thinking token.
- Developer Tools
Nub Bundles a Bun-Style Toolkit Onto Node Without the Runtime Swap
Nub ships a Bun-style all-in-one Node toolkit while keeping stock V8, trading Bun's native-module breakage for Node's own startup floor in a single Rust binary.
- Open Source
Bot-Account Lookups Miss 97% of AI Coding Agent Commits, 180M-Repo Census Finds
A 180-million-repo census finds bot-account lookups miss 97% of Claude Code commits, a 30x recall gap that makes prior AI-agent adoption estimates a floor.
- Security
How Reliable Are the LLM Judges Scoring Jailbreak Attacks?
Published jailbreak attack-success rates depend on the judges scoring them. A June 2026 audit finds both judge families miscalibrated and manipulable without removing harm.
- Models & Research
PV-TAM Corrects Decoding Drift and Boundary-Marker Bias in VLM Localization Scoring
PV-TAM, proposed in a June 2026 arXiv preprint, corrects two structural biases in VLM localization scoring: decoding drift and modality boundary-marker contamination.
- Agents & Frameworks
Do AGENTS.md Files Actually Help Coding Agents? A New Benchmark Tests It
An ETH Zurich benchmark finds AGENTS.md files don't lift task success rates but add over 20% to inference cost. A second study measures runtime savings, not success rate.
- Agents & Frameworks
Should AI Shopping Agents Pay Micro-Transactions for Verified Product Data?
A402 binds agent micropayments to verified delivery, but its TEE attests execution rather than product truth, leaving marketplaces to decide who refunds a wrong attribute.
- Models & Research
Meituan's General 365 Benchmark: Top Models All Score Under 63%
Meituan's General 365 benchmark caps Gemini 3 Pro at 62.8% and most of 26 models below 60%, exposing how saturated public leaderboards inflate reasoning scores.
- Models & Research
LLM Surrogates in A/B Tests: The 39% Recovery Gap and the Silent Bias Risk
A 2026 preprint proposes LLMs as A/B test proxies for human subjects. Raw outputs captured 39% of the treatment effect, and surrogacy bias does not average out.
- Models & Research
LLM Token Pricing vs Compute Cost: What the Tokenomics Math Shows
Token pricing decouples from GPU cost via cached discounts, reasoning markups, and per-model multipliers. No single per-token rate works as a budget unit for real workloads.
- Models & Research
Do LLM Judges Favor Their Own Output? A Sanity Check on Self-Preference
An LLM judge's self-preference claim holds only if the study fixed generation quality. Without that control, the bias number conflates real preference with a quality gap.