#llm
-
Stateless MCP: the protocol update that won back Simon Willison's interest
The July 28, 2026 revision of the Model Context Protocol did away with server-side sessions. Simon Willison explains why this simplifies development and makes agents safer — and shows off three tools of his own.
-
Opus 5 Is More Resistant to Prompt Injection
Anthropic's new model posted the best result on the IPI benchmark, outperforming both previous versions of Claude and all of its competitors. But it's still impossible to fully defend against prompt injections.
-
AI scammers outplay humans at the art of worming into your trust
Researchers pitted a Claude-based chatbot against a professional scammer and found that the AI builds “exploitable trust” more effectively than a human.
-
Baidu Releases vLLM-Kunlun — a Plugin for Running LLMs on Kunlun XPU Accelerators
Baidu has open-sourced vLLM-Kunlun, a pluggable module that lets you run the popular vLLM inference engine on its own Kunlun XPU accelerators without patching vLLM's source code.
-
DeepSeek Releases V4-Flash in Public Beta with a Focus on Agents
The updated deepseek-v4-flash model outperforms the Pro preview on agentic benchmarks and gains support for the Responses API format along with Codex adaptation.
-
68% of Peer-Reviewed Papers Turned Out to Be AI Slop: A View from the Trenches
Two reviewers for NeurIPS, WACV and an ECCV workshop went through 22 submitted papers and uncovered fabricated citations and impostor authors on a massive scale. Two papers with fake names were still accepted for oral presentations.
-
The Worm in the Document: How an Attack Self-Propagates Through Copilot in Word
A researcher has shown that instructions hidden inside a document can make Copilot for Word quietly alter figures in reports and copy malicious code into new files, turning them into carriers of the infection. Over 144 days, Microsoft was still unable to close the vulnerability.
-
GCC Will Reject Any Legally Significant AI-Generated Contribution — Except Tests
The GCC steering committee has approved a policy barring “legally significant” code generated by language models from the compiler. The only exception is for test cases.
-
AI Learns to Break Ciphers: The CryptanalysisBench Benchmark Finds New Attacks
A new benchmark measures how well large language models can do mathematical cryptanalysis. Some models are already finding vulnerabilities not previously described in the literature.
-
PyTorch as a Reference Language: Separating Specification from Implementation
Edward Yang of the PyTorch team proposes seeing the framework as both a reference specification and an implementation language, with a separate verifier to check any divergence between the two.
-
Moonshot AI prepares to release Kimi K3 — the first open model in the 3T class
An announcement for Kimi K3 by Moonshot AI has appeared on Hugging Face — according to its developers, the first open model at the three-trillion-parameter level. The weights are expected to be published on July 27, 2026.
-
The Gray Market for Tokens: How LLM Access Gets Resold at a Discount
An investigation by Matt Lenhard describes the ecosystem that has grown up around reselling language-model tokens by pooling API keys — mostly in China.
-
Anthropic Cut 80% of Claude Code's System Prompt — and Nothing Broke
Anthropic engineer Tariq Shihipar explains how the company radically simplified context for the new Claude 5 generation of models: fewer hard rules, more trust in the model's judgment.
-
Engineering Management After Code Got Cheap
A director of engineering reflects on which of the “old rules” of managing teams collapsed along with the cost of writing code — and which, on the contrary, now matter more than ever.
-
Canadian MP Reads a Chatbot's Prompt Aloud During a Speech in Parliament
A member of the New Brunswick legislature failed to notice that he was reading out not just his speech, but also a language model's aside offering a “more natural phrasing.”
-
Debian Votes on the Fate of LLMs in the Project
Debian developers have begun discussing a general resolution that will determine whether large language models may be used when working on the project.
-
Codeberg vs. AI: Armin Ronacher on the Risks of the New Rule
Codeberg has banned hosting projects written predominantly by generative AI. Developer Armin Ronacher explains why this attempt to protect open source could end up working against the very community it aims to defend.
-
Codeberg declares war on “vibe coding”: two resolutions against LLMs
Members of Codeberg e. V. have voted to refuse to train AI on user data and to ban projects generated by large language models. The platform sees LLMs as a threat to the free-software ecosystem.
-
Google Unveils Gemini 3.6 Flash and a Cybersecurity Model, While Gemini 4 Is Already in Training
The company has released three new models, including the first version of Gemini built for cybersecurity tasks — but the promised Gemini 3.5 Pro is once again nowhere among them.
-
Open Models Scare OpenAI. Should the US Be Afraid of Them?
The success of the Chinese open-weight model Kimi K3 has ignited a debate over banning such systems in the US. Behind all the talk of safety lies a far more mundane problem: frontier labs are finding it ever harder to turn AI into a business.
- More